SFF Build: Cubeor Aski 2


I’m moving to a different country for an unspecified length of time, and have made the difficult decision to leave my behemoth of a desktop PC behind. Initially I was planning on just living off of my Framework 12th gen laptop, which is a very capable little machine, but then managed to convince myself that it would be fun to build a small form factor desktop that it would be practical to ship to the other side of the world. My design goals in researching cases were:

  • Must be small enough to easily ship internationally, but large enough to fit a performance CPU and a cooler that is at least ‘adequate’
  • Does not have to have enough space for a GPU, but it’s a bonus
  • Would ideally be pleasing to look at as a desk object, since storage space will be at a premium in my new location.

After looking around, I ended up settling on the Cubeor Aski 2, a 10L small form factor computer with fetching laser-cut wood panels. Key features that seemed attractive are the space for a large number of fans (to help cool a performance CPU), the ability to insert a GPU without a riser (simpler build process) and aesthetics (I love the wood panelling). After ordering the case, I then specced out the internals as follows:

  • CPU: Intel i9-13900K
  • Cooler: Noctua NH-U9S
  • Motherboard: Gigabyte B760I Aorus Pro DDR4
  • Memory: Corsair Vengeance 64GB DDR4 kit (CL16)
  • Storage: Samsung 970 EVO 1TB
  • PSU: EVGA SuperNOVA 850 GM
  • Additional fans:
    • 1x NF-A14 140mm front intake fan (top)
    • 1x NF-F12 120mm front intake fan (bottom)
    • 2x NF-A8 80mm rear exhaust fans

The Case

Quick Look

Here are some photos of the case more-or-less as it appears out of the box. This is the ‘cherry’ colourway for the side panels. Note that the apparent lightness of the panels will vary as some photos are exposure compensated for the black elements of the case.

Initial Observations

  • Shipping cost is pretty reasonable for international, but budget some time for delivery. I placed my order on July 10th, the case shipped on the 14th, and arrived on the 24th.
  • Case arrived fully assembled and nicely packed. There were styrofoam plates on the top and bottom, and U-shaped spacers on the face and rear to ensure all sides were stood off from the shipping box. The entire box itself was also wrapped in some foam packaging material.
  • This thing is tiny! Here’s the Aski 2 next to my existing PC, which is inside a Nanoxia Deep Silence 5, a case which is about 70L (compared to the Aski 2’s 10L).
  • Case panels are affixed with what look M3x6 black flanged button head screws, so you’ll need a 2mm Allen key. The power button also uses M3 screws, but slightly longer ones (M3x10). The power button can be easily removed if necessary to wrangle the PSU.
  • The case doesn’t come with any motherboard mounting screws, and my motherboard only came with whatever thread screws come with ordinary ATX cases. The standoffs on the Aski 2 are M3 thread, though the thread starts surprisingly far down the standoff - there’s about 4mm of unthreaded lead-in. Not sure what purpose this serves, but I had some M3x12 button heads lying around, so I used 4 of these to secure the motherboard. The top of the standoffs are about 9mm from the top face of the internal metal panel, which leaves a little bit of room on the back for hiding cables.
  • The internal metal structure feels very sturdy. No flex at all. Riveted construction.
  • I got the extended feet, which are 10mm of rubber with a 10mm M4 threaded screw. They’re attached with a washer and ordinary nut; I’m personally a lock nut person but I doubt these are going anywhere. The screw is a bit longer than necessary (10mm is just a standard size), but it’s not like they’re going to interfere with a GPU mount. You could get different feet off McMaster if you really wanted to (e.g. something like this, so you just have the head of your screw in the case instead of the thread), but the durometer on the provided ones feels pretty good for a PC foot.

Foot thread extends into case slightly more than needed

  • The power button assembly is very cute. Comes off easily with two M3x10 screws. The physical actuator is 3d printed, but has a reasonably nice finish. We’ll see how it holds up; the fact that I can see the internal structure does make me worry it might delaminate or something with heavy use. That said, this is a button that’s getting pressed on the order of once a day. The switch itself is a tactile momentary switch that clips into the larger U-channel metal piece and then seems to be sandwiched by the smaller. Unfortunately this second piece is riveted in place, so replacing the tactile switch itself would be a little inconvenient, but again, very low duty part. The wood finish plate seems to be a press fit over the larger U-channel. Feels nice and snug now, though it’s a humid Boston summer, so possible that in a dry winter it may have more play with the lower humidity.
  • Taking the wood panels off, you can see that they’ve all been cleanly labelled with which panel it is and which way is up. This is actually important, because the side panels aren’t identical - they will not line up perfectly if installed backwards (see photos). When correctly installed, the panels have very nice tight clearances to each other. The panels also have little half-lasered holes to avoid interference with the rivets, which is a lovely touch. You’ll also see in the first photo below that the screws do leave little marks on the wood, but this is obviously covered by the screw head when installed.
  • One downside to this case is that it doesn’t come with any fan filters - luckily it ended up being fairly straightforward to add some in after the fact, but some sort of integrated solution might be nice for those that live with pets.
  • The rear 80mm fan mounting holes are a little tight - not sure if it’s just the powdercoat, but attaching the rear fans generated a fair few metal chips that people might want to keep an eye out for, especially if mounting fans after mounting the motherboard. In addition, the two sets of mounting holes are quite close together - so close that the silicone dampers on Noctua fans will actually interfere slightly. On the case-side of the fan, the mounting tension is enough to pull them in line most of the way, but I had to remove them from the internal side of the fan since otherwise they would not seat nicely.
  • Possibly a user error, but the front fan rail did deform somewhat after snugly installing a 140mm fan in the top front of the case. Possibly an adversarial positioning since this results in the screw being right in the middle of the cutout, rather than at one of the more supported ends.

Mount deformation on snug fan install

  • This isn’t a flaw of the case per se, but be warned that the tolerances for front fans with a GPU mounted is extremely tight! Even with one of the smallest GPUs I could find, the fit was so close I had to remove the silicone dampers on the fan in order to gain the extra fraction of a mm needed for it to fit cleanly. Looking at it the other way though, this case is exactly as big as it needs to be in this dimension.

Dust filters

As someone who lives around a lot of pet hair and general dander, dust filters are pretty crucial on any computer equipment to at least try and keep the inside clean. Unfortunately the Aski 2 doesn’t come with any, but I was able to cobble some reasonable ones together using some parts off McMaster (cheaper sources almost certainly exist). Namely,

  • A one foot square of 132” Flexible Magnet
  • A couple feet of 30x30 Nylon Mesh
    • If doing this again I might opt for a smaller mesh, though you still want it to be fairly open so you don’t restrict flow. The 61x61 mesh on McM is also a 50%+ open area option, so that’s probably a good bet.

The nylon mesh I purchased had no colourant, which made it a little conspicuous as a filter material out of the box. This was especially true for the PSU intake, which doesn’t completely fill the top panel (and if I made it oversized, it would have to have a cutout for the power button). To make the mesh blend in better, the first thing I did was dye it black. Dyeing nylon is slightly involved in that you need to use an acid dye. I opted for the Jet Black (639) from Jacquard dyes, which was readily available, with citric acid to achieve the necessary pH. Using a cheap baking tray as a vessel to avoid contaminating food use containers (if you do this, ensure the tray does not have a non-stick coating that will off-gas hazardous chemicals if heated directly) I dissolved the dye, soaked some pre-cut sheets of mesh in the dye solution, added citric acid solution, and then brought to a boil and let sit for thirty minutes while agitating. Note that getting the temperature up to boiling is critical for the dyeing process to work - I had to cover the pan with a metal sheet as a lid to keep heat from escaping.

With the mesh dyed, the previously visible edge of the filter material now blends pretty seamlessly with the dark powder coat on the chassis.

Next, I cut the magnetic sheet into 1cm wide strips using a rotary fabric cutter, and then slipped them onto the bare case chassis around the apertures I intended to put dust covers on (all of the ones that will be intakes). I then cut them down to size lengthwise by eye (cut one then layer it on the other to ensure they are consistent) then used them as a yardstick to measure out the nylon mesh, which I again cut using the rotary fabric cutter. Conveniently the mesh has a built in grid to make sure you cut it square; I found that cutting it slightly large then intentionally peeling off one edge strip at a time allowed fine tuning the size easily, after which you can use a ruler and the cutter to trim the loose ends. I also recommend sizing the mesh patch slightly narrower than the full width of the case, just so that you have some wiggle room to pull it taut over the apertures with the magnets. The adhesive on the back of the magnetic sheet I purchased is strong enough to tack the filter mesh in place, but unfortunately isn’t enough to keep it there long term, especially for the bottom intake which has gravity working against it. To reinforce the bond, I lay a bead of black RTV silicone adhesive (Permatex 81158) along the top of each interface and smoothed it into the mesh with a playing card. Once cured, any excess silicone was trimmed with a pair of scissors. With the additional adhesive, the filters now feel like they should hold up pretty well to light abuse. Note that on the front of the case, the combined thickness of the fan screws and magnet is too much for the front panel to seat comfortably. To get around this, I notched out the filter on that side to ensure a clean fit. Likewise, on the bottom intake, I notched the filter in around the four feet, and on the top intake narrowed one edge so that it can extend between the two attachment points for the top panel.

I’m quite pleased by how the filters turned out, especially with the black dye. In addition to blocking dust, the dark filters also help to hide the fans behind the front grille, leading to a nice clean look for the front of the case.

Build Process

Small cases generally require a fairly specific installation order for things to fit, and this one is no exception. I found it worked best to go in this order:

  1. Assemble the CPU, cooler, memory and SSD to the motherboard.
  2. Fit the motherboard into the case. I used M3x12 button head screws with nylon washers.
  3. Attach the auxiliary CPU power cable to the motherboard while it is still accessible.
  4. Install the GPU to the motherboard.
  5. Connect the power button to the motherboard. I chose to route it along the back side of the motherboard.
  6. Install the rear exhaust 80mm fans. I routed the cables on the top side, away from the motherboard. Both cables were connected to a splitter and then to the motherboard.
  7. Install the power supply, plug in the other side of the aux CPU power cable, and install the primary ATX power cable.
  8. Install the front fans. I found that I had to remove the silicone dampers in order to gain the extra mm of clearance needed to get them past the GPU.
  9. Install the GPU aux power cable. I chose to clip off the daisy chained connector for a cleaner fit. Ensure that if you do this you trim it very close to the connector to minimize risk of shorts (shorter than the example image).
  10. Cable manage the GPU aux power and rear fan cables along the back of the case using the intake mounting holes of the 80mm fans as binding posts.

System Performance

This case is about 1/7th the volume of my old PC, and I was curious to see how much raw CPU performance I was giving up, as well as just whether the very small noctua cooler would be able to keep up with the amount of power thrown off by the i9-13900K. To generate a realistic load for the work I usually do, I used a compile of the 6.1 Linux kernel as a benchmark. The system was initially run ouside the case on my desk, with unrestricted airflow, and only the onboard CPU cooler fan for air movement. Here are the results for that, along with reference results from my old desktop and my current laptop:

ATX Desktop (Threadripper 5965WX):   5m21.738
SFF Desktop (Intel i9-13900K):       6m14.078
Framework Laptop (i7-1260P):        20m17.199s

These results are honestly pretty shocking. The i9 is hot on the heels of the Threadripper, despite its $600 MSRP being only one quarter of the Threadripper’s $2400. I expected much more of a margin here. The Threadripper obviously has other advantages (ECC memory, more memory channels, dramatically more PCIe lanes, etc) but for this application it’s clear that the i9 blows it out of the water. Of course, both desktops completely outclass my laptop, which despite also being a modern CPU SKU can only hit about 35W of continuous package power with its fitted thermal solution.

Speaking of thermal solutions, the Noctua NH-U9S is rated by Noctua as ‘medium turbo/overclocking headroom’ for the i9-13900k, and I think that’s a pretty accurate assessment. The CPU very quickly hit 100°C package temp, and throttled the P-cores back from about 5.4GHz to about 4.9GHz by the end of the build, remaining at about 100°C the entire time. By way of comparison, the Threadripper with a NH-U14S TR4-SP3 sat at at (comparatively) cool 76°C and 4.3GHz all core clocks for the duration of the run.

Cooler Mod

To see if I could eke some more performance out of the NH-U9S, I browsed Digikey for alternate 92mm fans that might be able to push more air. The best replacement seemed to be the Delta AFC0912D-AF00, which offers much higher numbers than the Noctua across the board:

                    Noctua NF-A9        Delta AFC0912D-AF00
Airflow             1.315 m³/min        2.87 m³/min
Static Pressure     2.28 mmH₂0          13.35 mmH₂0
Acoustic Noise      22.8 db(A)          53.0 dB(A)
Power               1.2 W               9 W
Rated RPM           2000                4800

Note that that includes noise, and this fan is decidedly more of a ‘server-class’ sound than the Noctua fans. I crimped a standard PC fan header onto the Delta fan, attached it to the cooler tower using the Noctua mounting clips, and re-ran the compile:

                                    Time        Busy MHz    Package temp
SFF Desktop (With NF-A9 fan):       6m14.078    ~4990       100°C
SFF Desktop (With Delta fan):       5m59.175    ~5300        96°C

The additional airflow does actually manage to boost the P-core clocks by 300MHz or so, which certainly isn’t nothing. The package temp is also marginally lower, though still above the 85°C usually used as a cutoff for non-industrial silicon operating ranges. However, it comes at a significant noise cost. I don’t have a decibel meter to verify, but subjectively the fan is distractingly loud. 10dB is approximately a doubling of perceived loudness, so the thirty decibel rated difference between the fans suggests the Delta is 8x is loud as the Noctua. For the 4-5% increase in performance, this modification probably isn’t worth it given the design target of the PC.

Case Airflow

With a baseline for how the system performs in open air, I then installed it into the case along with all of the auxiliary fans and re-ran the benchmark. For these runs, the system is back to using the stock Noctua NF-A9 fan since that’s what I will likely run in the system going forward.

                            Time        Busy MHz    Package temp
SFF Desktop (Open Air):     6m14.078    ~4990       100°C
SFF Desktop (Inside Case):  6m06.618    ~5000       100°C

The result was actually a little faster, presumably because of the additional four fans channeling air around the system. The VRM on this motherboard is working pretty hard to feed the CPU, and has a relatively small heat sink that gets quite hot. Annoyingly, despite being present in the BIOS that temperature control doesn’t seem to be exposed by lm-sensors, but I’d wager it’s also happier with the guided case airflow, as are the memory DIMMs. At idle, all temperature zones measure around 40°C.


I’m very charmed by this case. I think it looks very attractive as a desk piece, is well built, and the internal design makes it surprisingly easy to work with despite the small size. Being able to fit a 2U PCIe card in there without a riser is also a very nice design achievement. The patterning on the front allows for pretty good airflow, even with the added dust filters. If I were to complain about anything, it would be that there are no filters included, but considering the space to work with it’s understandable. The complete lack of any front panel I/O could also be a downside depending on your preferences, but at least for me I’m alright trading it for a clean design.

DEFCON 29 Hardware Hacking Village CTF

This post contains writeups and some code samples for solving each of the puzzles from the DC29 HHV Challenge roughly organized by complexity. Many thanks to the HHV members that ran the CTF (rehr, wintermute, diogt and any others)!

Plenty of spoilers follow!


Backup Logs

For part a, we need to take the long view. Shrinking the interface size on Logic makes this a bit easier:

The Big Picture

For part b, careful examination of the clock signals (or blindly adding uart decoders to every signal) reveals that one of our channels actually transmits UART data for part of the trace.

Secret board

The filenames in the zip are all .gbr, or Gerber files for PCB fabrication data. Opening them in a viewer (e.g. gerbview or gerbv) reveals a nice little badge

What a nice badge

But what’s this, on the inner copper layers?

Inner Copper 1

Inner Copper 2

Circuit Cave

This is the way

I’d never seen a .circ file, and wasn’t sure exactly how to open it, so I just took a peek at the start to see what sort of format we were dealing with. Luckily, it comes with a clue!

ross@mjolnir:/h/r/Downloads$ head challenge.circ
<?xml version="1.0" encoding="UTF-8" standalone="no"?>
<project source="2.7.1" version="1.0">
  This file is intended to be loaded by Logisim (http://www.cburch.com/logisim/).

  <lib desc="#Wiring" name="0"/>
  <lib desc="#Gates" name="1"/>
  <lib desc="#Plexers" name="2"/>
  <lib desc="#Arithmetic" name="3"/>
  <lib desc="#Memory" name="4">
    <tool name="ROM">

After installing logisim from apt, we are greeted with a circuit that will shuffle out some obfuscated data on some seven segment displays. Presumably, this data is ascii for our flag. Writing down all the numbers by hand seems burdensome, but if you open the Similation -> Logging window you can add the output of the decoder circuit, and change the radix to 16 to record the hex value every time it changes. Combine this with the log to file option and a method to convert hex as ascii, and you have your flag. Or… close to your flag. It seems that we only actually want every 8th output digit, or whatever is loaded when the Counter(240, 350) signal wraps to zero (when the decimal point is lit).

To speed things up, I recommend changing the Simulate -> Tick Frequency value to something zippier than the default 1Hz.

No, this is the way

I’m a big fan of verilator! To get the flag here, I made one small addition to the tick code in the testbench:

  tb->i_clk = 0;

  // If the module has flagged the output data is valid, print that
  // character to stdout
  if (tb->o_out) {
    printf("%c", tb->o_data);

  if (tfp) {
    tfp->dump(tickcount * 10 + 5);

The only true way

This stage is similar, but uses nmigen instead of verilog. I’ve seen some nmigen stuff before, but haven’t stuck much of a toe in since I feel I need to thoroughly understand Verilog first. I couldn’t immediately spot a simple way to log data during the similation, so I did this the dumb way - opened up the trace in gtkwave and manually transcribed the data. Sometimes simple works!

GTKwave to the rescue

Biggest problem I had here was getting nmigen to run properly - there seemed to be some sort of dependency ordering issue, where if I installed nmigen and then nmigen_soc, it would downgrade nmigen to a version that doesn’t have similation. Installing nmigen_soc, then reinstalling nmigen again seemed to fix the problem for my virualenv.

Serial Swamp

Debuggin Interface

Looks like UART at first glance. Add a decoder with a nice default 9600 baud and we see a cute dinosaur in our console output:


Near the end we see a note that we are switching to fast uart to continue. Add a second decoder with the mildly zippier 115200 baudrate and we can see the rest of our boot prompt, including flag:

Pterodactyl dome

Lost Record

Wasn’t sure immediately what to make of this one, but googling our signal names (BCK, LRCLK, DOUT) suggests that this is an i2s (audio) interface with 2 channels. Counting our clock cycles, it looks like we have 32 bits per sample, and scanning through some of our early data samples it seems likely that the values are signed, encoded as 2’s complement (instead of toggling between min and max amplitude very quickly).

First step here is to reconstruct the audio, so let’s add an i2s analyzer in Logic and export the data as a CSV. We can then use python to scarf the data and emit a flat binary file with packed sample data, like so:

import csv
import sys
import struct

class Sample(object):
    def __init__(self, time, channel, data):
        self.time = time
        self.channel = channel
        self.data = data

# Read the CSV data
samples = []
with open(sys.argv[1]) as f:
    reader = csv.reader(f)
    reader.next()  # Skip header
    for row in reader:
                float(row[2]), # Timestamp
                int(row[5]),   # Channel (L/R)
                int(row[6])    # Sample data
        except Exception, e:
            print e
            print row

# Calculate sample rate
time_start = samples[0].time
time_end = samples[-1].time
sample_time = time_end - time_start
sample_count = len(samples)
samples_per_sec = sample_count / sample_time / 2
print "Sample rate: %f" % samples_per_sec

# Create an array of data for each channel
ch0 = []
ch1 = []
for sample in samples:
    if sample.channel == 0:
    if sample.channel == 1:

# Interleave the data and pack down into a file
with open('out.stereo.bin', 'wb') as f:
    for pair in zip(ch0, ch1):
        f.write(struct.pack('ii', pair[0], pair[1]))

We can test that the audio sounds right by playing it raw:

aplay -f S32_LE -c 2 -r 44100 out.stereo.bin

My god! They’re coming back! But that doesn’t seem to be the trick. Let’s load this audio into audacity so that we can poke around a bit -

File -> Import -> Raw Data, Signed32LE, 2 channels, 44100 Hz

A fun steganographic technique with audio files is embedding pictures in the Spectrogram of the track. Let’s enable spectrogram view and zoom on in to that final segment:

Suggested settings for spectrogram view

Who drew this here

Lab Control

This one took some doing, mostly due to the sheer size of the datasheet for this chip, and the lack of a clear transaction process overview. But, first things first, we need to make sense of this data - scanning the trace in Logic, it looks like some bog standard SPI transactions. From the chip datasheet, it looks like spi transactions consist of an address and Read/~Write bit, followed by the read/written data. So, let’s decode the data in Logic and export a CSV we can run through some analysis. The output of our export will look like the following:


First things first is to load this data into a programming language where we can mess around a bit, so here’s a quick and dirty loader in C++:

// A 'transaction' begins when ~CS goes low, ends when it goes high.
// Multiple bytes could be transferred per transaction.
struct Txn {
  std::vector<uint8_t> mosi;
  std::vector<uint8_t> miso;

// ...

// Load input file
const char *filename = argv[1];
std::fstream fstream_in;
fstream_in.open(filename, std::ios::in);
if (!fstream_in.is_open()) {
  fprintf(stderr, "Failed to open %s\n", filename);
  return -1;

// Split each line, and extract our important fields
std::string line;
std::vector<Txn> txns;
bool is_in_txn = false;
Txn active_txn;
while (std::getline(fstream_in, line)) {
  std::stringstream row_ss(line);
  std::string field;
  std::vector<std::string> fields;
  while (std::getline(row_ss, field, ',')) {
  if (fields[1] == "\"enable\"") {
    is_in_txn = true;
  } else if (fields[1] == "\"result\"") {
    active_txn.mosi.emplace_back(strtoll(fields[4].c_str(), nullptr, 16));
    active_txn.miso.emplace_back(strtoll(fields[5].c_str(), nullptr, 16));
  } else if (fields[1] == "\"disable\"") {
    active_txn = {};
  } else {
    fprintf(stderr, "Unknown op '%s'\n", fields[1].c_str());

Once we have all the transactions, each of which starts with an address, we can scan through for addresses that are ‘interesting’. First one that looks promising to me is FIFODataReg - if there’s data, it’s going to be passing through there. So what happens if we print all the contiguous reads/writes to that register?

bool last_was_fifo_read = false;
for (auto &txn : txns) {
  // MSB controls read/write
  const bool is_read = txn.mosi[0] & (1 << 7);
  // Address is left shifted one bit just to trip you up
  const uint8_t addr = (txn.mosi[0] & 0x7F) >> 1;

  // Contiguous read from FIFO?
  if (is_read && addr == 0x09) {
    if (!last_was_fifo_read) {
      printf("\n[RD] ");
    printf("%02x", txn.miso[1]);
    last_was_fifo_read = true;
  } else {
    if (last_was_fifo_read) {
    last_was_fifo_read = false;

  // Contiguous write to FIFO?
  // [...]

If we run this, we see some interesting long strings of hex data:

[WR] 9320
[WR] 26
[RD] 0400
[WR] 9320
[RD] 8634cd1f60
[WR] 93708634cd1f60
[WR] 93708634cd1f60f5d6
[RD] 08b6dd
[WR] 6008f02d53fda4a78634cd1f
[WR] 3008
[WR] 30084a24
[RD] 6178656c5f3532ffffffffffffffffff
[WR] 6016c47650e887a08634cd1f
[WR] 3016
[WR] 3016b5dd
[RD] 456c65637461646f6e6c6f67697374ff
[WR] 6022f22ed77f55768634cd1f
[WR] 3022
[WR] 302212aa
[RD] 6868765f6c616273ffffffffffffffff
[WR] 26

To get a bit more context, let’s also print some of the other registers that are accessed:

std::unordered_map<uint8_t, std::string> reg_names;
reg_names[0x00] = "Reserved";
reg_names[0x01] = "CommandReg";
reg_names[0x02] = "ComlEnReg";
reg_names[0x03] = "DivlEnReg";
reg_names[0x04] = "ComIrqReg";
reg_names[0x05] = "DivIrqReg";
reg_names[0x06] = "ErrorReg";
reg_names[0x07] = "Status1Reg";
reg_names[0x08] = "Status2Reg";
reg_names[0x09] = "FIFODataReg";
reg_names[0x0a] = "FIFOLevelReg";
reg_names[0x0b] = "WaterLevelReg";

for (auto &txn : txns) { // Same loop as prev prints

  std::string reg_name = std::to_string(addr);
  if (reg_names.find(addr) != reg_names.end()) {
    reg_name = reg_names[addr];
  const uint8_t datum = is_read? txn.miso[1] : txn.mosi[1];
  fprintf(stderr, "%s to %s: %02x", is_read ? "READ" : "WRITE", reg_name.c_str(), datum);

If we re-run, we now have some control register writes happening right after our fifo transaction:

[WR] 6022f22ed77f55768634cd1f
WRITE to CommandReg: 0e

Now we’re getting somewhere - right after we write all this data to the FIFO, we’re doing something with a CommandReg. What does 0x0e correspond to?

    if (!is_read && addr == 0x01) {
      switch (txn.mosi[1]) {
      case 0: printf("Command: Idle\n"); break;
      case 3: printf("Command: CalcCRC\n"); break;
      case 4: printf("Command: Transmit\n"); break;
      case 8: printf("Command: Receive\n"); break;
      case 12: printf("Command: Transceive\n"); break;
      case 14: printf("Command: MFAuthent\n"); break;
[WR] 6008f02d53fda4a78634cd1f
Command: MFAuthent

MFAuthent - if we search the datasheet for this, we get a hit well over a hundred pages in:

MFAuthent packet structure

If we break down the data written to the FIFO right before this command, it looks like it matches nicely:

60              # Authentication command
08              # Block address
f02d53fda4a7    # Sector key
8634cd1f        # Card serial number (Uid)

If we combine the data from this write (UID and sector key) with the data from the Transceive calls that happen right after it, we have all the data we need for our key:

[WR] 30084a24
Command: Transceive

[RD] 6178656c5f3532ffffffffffffffffff
Command: Idle

Broken Display

This is a fun one. We have a trace with a clock, MOSI, reset and ‘DC’ line, and know that we are talking to a display controller. These controllers tend to have Data vs Control modes, so this is likely what our DC signal represents. In order to attack this, we will want to export the commands sent to the LCD, and then run them through a little simulator to recreate the pixel state that the display would have. First step is to get the data out of Logic and into something more parsable. To do this, I added two SPI decoders - both use CLK, MOSI and DC as the serial clock / data / chip select lines, but one of them has CS set as active low and the other has it active high. This way, one will only trigger for data mode transfers and the other only triggers for control mode transfers. If we name these decoders nicely and export, we get data like so:

ross@mjolnir:/h/ross$ head display.csv

with a little shell magic, we can preprocess this a little to make our lives easier later:

ross@mjolnir:/h/ross$ cat display.csv | grep result | cut -d , -f 1,5 > data_cmd.csv
ross@mjolnir:/h/ross$ head data_cmd.csv

Now, let’s load this up and see what’s involved in replaying this data. First, some boilerplate:

struct Packet {
  enum Type {
  Type type;
  uint8_t data;

// [...]

// Parse file
const char *filename = argv[1];
std::fstream fstream_in;
fstream_in.open(filename, std::ios::in);
if (!fstream_in.is_open()) {
  fprintf(stderr, "Failed to open %s\n", filename);
  return -1;
std::string line;
std::deque<Packet> packets;
while (std::getline(fstream_in, line)) {
  std::string::size_type comma_pos = line.find(",");
  std::string type = line.substr(0, comma_pos);
  std::string value_str = line.substr(comma_pos + 1);
  Packet p;
  p.type = type == "\"data\"" ? Packet::Type::DATA : Packet::Type::COMMAND;
  p.data = strtol(value_str.c_str(), nullptr, 16);

We now have a queue of command and data packets. From here, we just need to peel off the first command, see how many data bytes we need to associate with it, update the state of our fake LCD, and rinse and repeat. In fact, it turns out we can ignore most of the setup commands, though it’s useful to note that the LCD is programmed in 16-bit data mode, where each pixel is RGB565. Skipping through the rest of the commands from the datasheet, we come across three that we definitely want to implement - the Row/Column window select, and the RAMWR commands. The LCD controller in question does not allow you to addresss the entire memory space at once, instead you have to define an X region xs (x start) to xe (x end), and similarly a Y region, into which the RAMWR function will write. So, let’s do some accounting for these in a little loop:

// Cannot understand why STL containers don't include a method for this
auto next_packet = [&]() -> Packet {
  auto p = packets.front();
  return p;

// Row/column window, and current pixel write address
uint16_t xs = 0, xe = 0xef;
uint16_t ys = 0, ye = 0x13f;
uint16_t x_addr = xs;
uint16_t y_addr = ys;

// State machine
while (packets.size()) {
  // Take first packet
  auto packet = next_packet();

  // Ignore unexpected data packets
  if (packet.type == Packet::Type::DATA) {
    fprintf(stderr, "Unexpected data packet %02x\n", packet.data);

  // Switch next packet command
  switch (Command(packet.data)) {

  case Command::COL_ADDR_SET: {
    // Consume 4 data packets
    auto xs15_8 = next_packet();
    auto xs7_0 = next_packet();
    auto xe15_8 = next_packet();
    auto xe7_0 = next_packet();
    xs = xs15_8.data << 8 | xs7_0.data;
    xe = xe15_8.data << 8 | xe7_0.data;
    x_addr = xs;
    fprintf(stderr, "xs: %4d, xe: %4d\n", xs, xe);
  } break;

  case Command::ROW_ADDR_SET: {
    // Consume 4 data packets
    auto ys15_8 = next_packet();
    auto ys7_0 = next_packet();
    auto ye15_8 = next_packet();
    auto ye7_0 = next_packet();
    ys = ys15_8.data << 8 | ys7_0.data;
    ye = ye15_8.data << 8 | ye7_0.data;
    y_addr = ys;
    fprintf(stderr, "ys: %4d, ye: %4d\n", ys, ye);
  } break;

  default: {
    fprintf(stderr, "Unhandled command 0x%02x\n", packet.data);
  } break;

Now that we have the row/column data, we need to actually write pixels to those addresses. We create a buffer to hold our pixel data & some helper tools to write data, then handle that command type in our loop like we did the row/col addresses:

// Display is a known size
const int rows = 240;
const int cols = 320;
static uint8_t pixels[rows * cols * 3];
auto write_pixdata = [&](int row, int col, uint16_t data) {
  // input data is rgb 565
  uint8_t r = (data >> 11) & 0b11111;
  uint8_t g = (data >> 5) & 0b111111;
  uint8_t b = (data >> 0) & 0b11111;
  const int offset = (row * cols + col) * 3;
  pixels[offset + 0] = r;
  pixels[offset + 1] = g;
  pixels[offset + 2] = b;

// [...]

// Inside our switch statement from above
case Command::RAMWR: {
  // Consume as many data packets as we can
  int wrote = 0;
  while (packets.front().type == Packet::Type::DATA) {
    // Pop 2 data packets
    auto dp0 = next_packet();
    auto dp1 = next_packet();
    uint16_t dat = dp0.data << 8 | dp1.data;
    write_pixdata(x_addr, y_addr, dat);

    // Handle row/column address increment
    if (x_addr > xe) {
      x_addr = xs;
    if (y_addr > ye) {
      y_addr = ys;
  fprintf(stderr, "Wrote %d pixels\n", wrote);
} break;

This should be enough to regenerate a pixel buffer, but we need a way to see it. Luckily, openGL makes this relatively straightforward - we can create a texture, load the pixel data into it, then render it as a flat square using the following incantations:

// Glut needs a render functin to call; annoyingly can't use a std::function
// so work around by just making key state static
static int window_w = cols, window_h = rows;
static GLuint gl_texture;
auto gl_display = []() {

  // Generic 2d orthographic view
  glViewport(0, 0, window_w, window_h);
  glOrtho(0, window_w, window_h, 0, -1, +1);

  // Render texture as full screen
  glBindTexture(GL_TEXTURE_2D, gl_texture);
  glTexCoord2f(0.0f, 1.0f);
  glVertex2f(0, window_h);
  glTexCoord2f(0.0f, 0.0f);
  glVertex2f(0, 0);
  glTexCoord2f(1.0f, 1.0f);
  glVertex2f(window_w, window_h);
  glTexCoord2f(1.0f, 0.0f);
  glVertex2f(window_w, 0);
  glBindTexture(GL_TEXTURE_2D, 0);


auto gl_resize = [](int w, int h) {
  window_w = w;
  window_h = h;

// Initialize GLUT
glutInit(&argc, argv);

// Create a window to display in
glutInitWindowSize(window_w, window_h);
glutInitWindowPosition(0, 0);
glutCreateWindow("Broken Display");

// Set up our render callbacks

// Generate texture using our pixel data
glGenTextures(1, &gl_texture);
glBindTexture(GL_TEXTURE_2D, gl_texture);
glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB, cols, rows, 0, GL_RGB,
             GL_UNSIGNED_BYTE, pixels);
glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
glBindTexture(gl_texture, 0);


With this, we get a cute little QR code in hacker green. My phone had a hard time reading this as-is, so I had to screenshot it, open it in GIMP and value invert the colours to get something it was happier with.

Unbroken Display

FPGAs 3: Wishbone and Soft Cores

Companion code for this post available on Github

One of the core design patterns in the arsenal of an FPGA developer is the finite state machine. Such systems can be small, fast, easy to reason about and extremely powerful for sequential logic. But there can come a point where a state machine grows so complex that the hardware implementation starts to become extremely costly, or perhaps you want to be able to update the behaviour of a large state machine from an external memory. At such a time, it may make sense to consider replacing a complex special-purpose state machine with a highly evolved general purpose state machine - in other words, a CPU.

In this post, we will cover the basic idea of what a soft CPU is, how to connect a CPU to ROM and RAM using the Wishbone bus, how to write firmware for our custom SoC, how to use interrupts, and how to build our own memory-mapped IO peripherals. This post will assume some familiarity of Verilog and C++ code; if readers are less familiar with Verilog I would recommend starting with this post that attempts to explain what FPGAs are, and how Verilog can be used to command them.

Soft CPUs

If one considers the stripped down structure of a CPU, they can start to see how a CPU is, at its core, simply a state machine - it loads an instruction, decodes the instruction, performs some operation based on the instruction, increments the program counter and transitions back to the fetch state. One could definitely have some fun creating their own minimal CPU with a custom instruction set, but the level of complexity and tooling that goes into a performance-competitive CPU is no mean feat. Thus, for someone that wants to get a system built with minimal reinvention, it makes sense to utilize one of the increasingly many CPUs available online. There are many to choose from - this article has a good rundown of some free options, and even proprietary designs such as the ARM Cortex-M0 are available for evaluation. However, the complexity of integration and licensing headaches are somewhat of a negative for ARM cores in particular. For the purposes of this article, we will focus instead on the highly customizable and increasingly popular RISC-V ISA, which boasts multiple free implementations targeting different use cases.

The shortlist


One of the earliest such CPUs. Optimized primarily for small size and high max frequency, this core has weaker performance per clock cycle numbers than some other CPUs but is easy to integrate and can fit handily in even some of the smaller FPGAs on the market.


One of the more unique cores, SERV is a bit-serial architecture - by taking the tradeoff between clock cycles and design area to the extreme, this CPU executes on only one bit at a time (instead of on 32 bits at once), reducing the size of the core to the point that at least 16 cores may be instantiated on the ICE40LP8K, an FPGA with only 7680 logic elements.


Of the readily available FOSS RISC-V cores, the VexRiscv is certainly the most configurable. The core itself is written using Spinal HDL, a set of Scala libraries, which allows for higher level components to be tweaked more readily than would be possible in straight Verilog. It can be tuned all the way from a minimal 500 LUT core with no hardware multiply or interrupt support all the way up to a 3000 LUT variant with caches, interrupts, branch prediction and MMU that allows one to run a full Linux core.

Customizing the VexRiscv

Given the customizability of the VexRiscv, and the superior performance per clock cycle compared to the other options, we will use it as our base for this project. The downside to customizability is understanding all these configuration options, so the first thing to do is have a read through some of the example core generation files, and write our own CPU definition to meet our needs.

If we take a look at the demo folder in the VexRiscv repo, we see a number of templates we can use to base our core off. We will mostly pull from the GenFull example, and our full custom configuration and generated CPU file can be found in the repo for this blog post here.

If we read through our custom GenVexRiscv.scala, we can see that the CPU is constructed of various plugins. Most are named in a fairly self-explanatory way, or do not require much modification, so we will touch on only some of them directly here. The first is our ibus, or instruction data bus plugin:

// We need an instruction data bus on the CPU. This bus is separate
// from the data bus for performance reasons, and here we will
// instantiate the cached version of this plugin, which is a
// significant performance improvement on a non-cached implementation
new IBusCachedPlugin(
  // We want to be able to set the reset address in verilog later, so
  // leave it null here
  resetVector = null,
  // Conditional branches are speculatively executed.
  // There is no tracking of whether a branch is more likely to be
  // executed or not
  prediction = STATIC,
  // Include a 4KiB instruction cache
  config = InstructionCacheConfig(
    cacheSize = 4096,
    bytePerLine = 32,
    wayCount = 1,
    addressWidth = 32,
    cpuDataWidth = 32,
    memDataWidth = 32,
    catchIllegalAccess = true,
    catchAccessFault = true,
    asyncTagMemory = false,
    twoCycleRam = true

Caching is one of the most powerful performance tools there is for CPUs, so we want to be sure to add a generous cache to our CPU here. Otherwise, non-cached instruction fetches will require us to spend at least 2 cycles asking for data over the wishbone bus. If you have to wait two cycles for each new instruction, your effective clock speed has already been cut in half! If your code is stored on external memories, such as a SPI flash, your uncached performance will be even worse as you have to potentially spend many clocks sending addresses and data back and forth over a SPI interface for each new instruction.

The next section that we want to pay some special attention to is the CsrPlugin. The RISC-V ISA defines a number of configuration and status registers, but it is not necessarily required that all are present, readable or writable in a given implementation. We will want to make our CPU flexible when it comes to interrupt configuration, and we would also like to be able to use the cycle counter register, so we will configure registers mtvec and mtcycle with READ_WRITE access. We could also set some CPU identification registers here, if for example our firmware would run on different flavours of soft CPU and would need to determine capabilities at runtime.

// Implementation of the Control and Status Registers.
// We want to make sure that registers we use for interrupts, such as
// mtvec and mcause, are accessible. We have also enabled mcycle
// access for performance timing.
new CsrPlugin(
  config = CsrPluginConfig(
      catchIllegalAccess = false,
      mvendorid      = null,
      marchid        = null,
      mimpid         = null,
      mhartid        = null,
      misaExtensionsInit = 66,
      misaAccess     = CsrAccess.NONE,
      mtvecAccess    = CsrAccess.READ_WRITE,
      mtvecInit      = 0x80000000l,
      xtvecModeGen   = true,
      mepcAccess     = CsrAccess.READ_WRITE,
      mscratchGen    = false,
      mcauseAccess   = CsrAccess.READ_ONLY,
      mbadaddrAccess = CsrAccess.READ_ONLY,
      mcycleAccess   = CsrAccess.READ_WRITE,
      minstretAccess = CsrAccess.NONE,
      ecallGen       = false,
      wfiGenAsWait   = false,
      ucycleAccess   = CsrAccess.READ_ONLY,
      uinstretAccess = CsrAccess.NONE

The final thing that we will do to this CPU is ensure that it speaks the Wishbone protocol on the instruction and data buses. Wishbone is a standard protocol for on-chip communication, and has the benefit of being very simple to implement. Luckily the VexRiscv comes with a built-in function to transform the interface to wishbone, so we simply need to invoke it:

// CPU modifications to use a wishbone interface
cpu.rework {
  for (plugin <- cpuConfig.plugins) plugin match {
    case plugin: IBusSimplePlugin => {
    case plugin: IBusCachedPlugin => {
    case plugin: DBusSimplePlugin => {
    case plugin: DBusCachedPlugin => {
    case _ =>

When we are happy with our configuration file, we need to generate a Verilog output we can feed to out synthesis tools. To do so, first clone the VexRiscv repo, and install the Scala build tool sbt. Then add our generation script, and build:

# Clone the VexRiscv repo
git clone https://github.com/SpinalHDL/VexRiscv.git
# Ensure we have the scala build tool
sudo apt install sbt
# Clone the associated code for this blog post
git clone https://github.com/rschlaikjer/fpga-3-softcores.git
# Copy our core generation spec into the VexRiscv repo
cp fpga-3-softcores/vendor/vexriscv/GenVexRiscv.scala VexRiscv/src/main/scala/vexriscv/demo/
# Move into the Vexriscv repo, and buld our core
cd VexRiscv
sbt "runMain vexriscv.demo.GenVexRiscv"

The first invocation of sbt may take some time as it resolves dependencies, but at the end you should end up with a new VexRiscv.v at the top of the repo. Step one is complete - we have our very own CPU.


Wishbone is a simple on-chip logic bus, with relatively low complexity required to implement the ‘classic’ non-pipelined interface. Other buses, such as AXI from the AMBA family of interconnects from ARM, are more powerful but come with a higher cost to implement, both in complexity and logic area. For our application, wishbone is more than sufficient.

The classic Wishbone cycle

The basic Wishbone interface consists of the following signals, some of which travel from the master to the slave (m2s) and some of which travel from the slave to the master (s2m):

wire [31:0] m2s_adr; // Address select
wire [31:0] m2s_dat; // Data from master to slave (for writes)
wire  [3:0] m2s_sel; // Byte select lanes for write enable
wire        m2s_we;  // Write enable (active high)
wire        m2s_cyc; // Cycle in progress. Asserted high for the duration of
                     // the transaction
wire        m2s_stb; // Strobe output. Asserted high to indicate data is valid
                     // for transfer from the master to the slave
wire [31:0] s2m_dat; // Data from slave to master (for reads)
wire        s2m_ack; // Read data valid strobe

The basics of a wishbone transaction are that the master must assert the m2s_cyc line to indicate that a transaction cycle is in progress. On the same cycle, or any subsequent cycle before m2s_cyc is deasserted, the master may load the address, write data, write enable and write select signals and asssert the m2s_stb signal to indicate to the slave that these data are valid. The slave will then perform whatever operation has been requested (read or write as according to m2s_we), and when ready present output data (if necessary) on the s2m_dat signal and strobe the s2m_ack signal to indicate that the data is valid. The master may then initiate another operation by asserting m2s_stb, or release m2s_cyc to end the transaction.

As an example, here is a trace of a wishbone transaction. At time marker A, the master outputs the valid address, data and control lines and simultaneously asserts the cyc and stb lines. One the next positive edge of the wishbone clock wb_clk, we see the slave respond with new data on its output and the assertion of the ack signal. At time marker B, the master clears the cycle and strobe signals, ending the transaction.

Classic Wishbone Transaction

There are other signals specified in the full Wishbone spec, such as errors, retry signals and tag data, but for many simple peripherals these signals are not implemented. For complex designs beyond the one shown here, or for further reference on the signals or formal properties of the bus, the full Wishbone Spec is an excellent reference. But for now, we know enough to be dangerous, and create some peripherals for our CPU.

The persistence of memory

To make a soft CPU useful, it needs to be fed instructions from somewhere. It’s also fairly common to want some amount of random access memory in excess of the built-in registers. To solve both of these problems, let’s write a quick memory module that can be accessed over the wishbone bus to provide either instructions or data to our CPU. Since both ROM and RAM are similar in behaviour here, we can reuse the same design for both. The implementation of a minimal working memory is something like the following (full code):

// Our main data storage. Yosys will replace this with an appropriate block ram
// resource.
parameter SIZE = 512; // In 32-bit words
reg [31:0] data [SIZE];

// Each data entry is 32 bits wide, so right shift the input address
// Sub-word indexing is achived through the byte select lines
localparam addr_width = $clog2(SIZE);
wire [addr_width-1:0] data_addr = i_wb_adr[addr_width+1:2];

integer i;
always @(posedge i_clk) begin
    if (i_reset) begin
        // Under reset we can't zero the memory contents, just ensure
        // that the wishbone bus returns to an idle state.
        o_wb_ack <= 1'b0;
    end else begin
        // Always ensure that our ack strobe defaults to low
        o_wb_ack <= 1'b0;

        // If we are being addressed, and this is not the cycle following an
        // ack, we need to perform a read/write
        if (i_wb_cyc & i_wb_stb & ~o_wb_ack) begin
            // We always read/write in the same cycle, so can
            // unconditionally set the ack here
            o_wb_ack <= 1'b1;

            // Always read the data at the given address to the output register
            // If this was a write operation, the data will simply be ignored
            o_wb_dat <= data[data_addr];

            // If this is a write, we need to do something a bit special here.
            // Since this memory needs to support non-32-bit operations, we
            // have to respect the byte select values. Handle this by looping
            // over each bit, and only writing the corresponding word if the
            // select bit is set.
            if (i_wb_we) begin
                for (i = 0; i < 4; i++) begin
                    if (i_wb_sel[i])
                        data[data_addr][i*8 +: 8] <= i_wb_dat[i*8 +: 8];

With this, we have something that can be read and written over wishbone (and we can even add some test cases for it). This looks good for RAM, but for ROM we need a way to preload our firmware into the memory. Verilog has system functions, $readmemb and $readmemh that can load binary or hex data into a memory from a file. So, let’s add another small feature so that we can tell our ROM memory to initialize itself from the firmware hex we will generate later:

// If we have been given an initial file parameter, load that
parameter INITIAL_HEX = "";
initial begin
    if (INITIAL_HEX != "")
        $readmemh(INITIAL_HEX, data);

Connecting the blocks

Now that we have a CPU with a wisbone interface, and two wishbone peripherals, we just need to connect them all together. Writing all of the muxing and switching logic for this by hand would be extremely tedious, but luckily there exists an excellent tool, wb-intercon, that does the legwork for us. Given a YAML description of a bus (the masters, which slaves they connect to and what address they should see those slaves at) wb-intercon can generate all of the verilog necessary to mux and arbitrate the various signals.

So, to start off with, let’s create a very simple wishbone layout that has two masters (our CPU instruction and data bus ports) and two slaves, our ROM and RAM blocks. Note that our ibus here only connects to the ROM, so executing from memory isn’t allowed, but could of course be changed if you wanted to! We will arbitrarily locate the rom at 0x20000000, and give it a max size of 4KiB. We will then locate the RAM at 0x80000000, and only give ourselves 2KiB.

      slaves: [cpu0_rom]

      slaves: [cpu0_rom, cpu0_ram]

      offset: 0x20000000
      size:   0x1000

      offset: 0x80000000
      size:   0x800

If we run this through the wb-intercon generator, which in the full project is done as part of the CMake build, we end up with two files - an implementation file, which contains the modules responsible for the address decoding and signal multiplexing, and a header file that contains all of the signal definitions for our various bus participants. In order to connect up our modules, in our top gateware file we just need to include this header and then use the generated wires to connect up our two wishbone RAM blocks, like so:

// Include the header that defines all the wishbone net names
`include "gen/wb_intercon.vh"

// We initialize this directly from the hex file with our firmware in it
wb_ram #(
    .SIZE(1024), // In 32-bit words, so 4KiB
) cpu0_rom (

wb_ram #(
    .SIZE(512) // 2KiB
) cpu0_ram (

Once we similarly connect the iBusWishbone and dBusWishbone signals on our CPU, everything is in place for us to start writing some code to run on our new system.

Baremetal RISC-V programming

Now that we have defined our hardware, we need to start defining our firmware. Since we do not have any operating system to handle hardware initialization and program startup for us, we must do it ourselves. This means getting hands on with the linker, the assembler and some low level features of the RISC-V architecture.

Defining our layout

At the end of our firmware compilation process, we need to end up with a series of bytes that represent the code our CPU should run and the values of any initialized data. We can then load these bytes into the CPU, and should be up and running. However to get here, we need to start by defining for the compiler how to arrange that code, and at what addresses the running program can expect to find things. In our earlier wishbone intercon YAML, we specified that the ROM should appear at 0x20000000, and the RAM at 0x80000000. Since we want to locate non-volatile data in ROM and volatile data in RAM, we need to tell the linker where these two memories exist, and what goes into each, so that it can generate correctly addressed loads, stores and jumps.

The way to do this is in the linker script, the complete version of which is here. We define our memory regions (in this case we have just 2) with a MEMORY directive, like so:

/* Define our main memory regions - we created two memory blocks, one to act as
 * RAM and one to contain our program (ROM). The address here should match the
 * address we gave the memories in our wishbone memory layout.
    ram (rwx)   :   ORIGIN = 0x80000000, LENGTH = 0x00000800
    rom (rx)    :   ORIGIN = 0x20000000, LENGTH = 0x00001000

This tells the linker that we have two regions where we can store data, and how much data we can safely fit in each location. Both these values must match the values used in the gatware, or there may be subtle problems later! Now that we have defined the regions, we need to indicate what parts of our program live in which section. We’ll start with the code itself, or the .text section - this should contain our initial startup code at the very beginning, followed by the rest of our code, and any other read-only data:

.text : {
    /* Ensure that our reset vector code is at the very beginning of ROM,
     * where our CPU will start execution
    /* General program code */
    /* Ensure that the next block is aligned to a 32-bit word boundary */
    . = ALIGN(4);
    /* Read-only data */
} >rom /* Locate this group inside the ROM memory */

Linker Relaxation

There are various other sections that must be located, all of which may be found in the full linkerscript, but the other section I will flag here is one specific to the RISC-V architecture - the .sdata (small data) and .sbss (small bss) sections.

The problem with a 32-bit ISA with 32-bit wide instructions is that it is impossible for a single instruction to encode an offset that can address the entire memory space - in reality, the immediate forms of memory load/store and jump instructions can address only up to 21 bits from the current PC / other base pointer. What this means is that operations referencing addresses over 2^21 bits away must be broken up into two operations - one to load an immediate into a register, and another to perform the actual operation. This can be a significant performance problem for some code, so RISC-V includes a feature called Linker Relaxation. When the linker is assembling the final binary, it will attempt to emit smaller jump/load instructions by addressing them relative to a special register, the global pointer (gp) register. This register is expected to be set to the address 0x800 bytes past the start of the small data section, such that it can be used as a base for single-instruction addressing. For a fuller explanation of linker relaxation, the SiFive blog has an excellent post here.

What makes this relevant to us is that we have some new sections in our linker script, that we might miss if we were to assume the same layout as some other embedded systems. In our linker, we need to be sure to actually include the .sdata and .sbss sections (otherwise, statically initialized variables may silently become zero-initialized!) and export the desired location of the global pointer. For our data section, this results in a linker directive like the following:

/* Our data segment is special in that it has both a location in rom (where
 * the data to be loaded into memory is stored) and in ram (where the data
 * must be copied to before main() is called).
.data : {
    /* Export a symbol for the start of the data section */
    _data = .;
    /* Insert our actual data */
    . = ALIGN(4);
    /* Insert the small data section at the end, so that it is close to the
     * small bss section at the start of the next segment
    __global_pointer$ = . + 0x800;
    . = ALIGN(4);
    /* And also make a note of where the section ends */
    _edata = .;
/* This section is special in having a Load Memory Address (LMA) that is
 * different from the Virtual Memory Address (VMA). When the program is
 * executing, it will expect the data in this section to be located at the
 * VMA (in this case, in RAM). But since we need this data to be
 * initialized, and RAM is volatile, it must have a different location for
 * the data to be loaded _from_, the LMA. In our case, the LMA is inside the
 * non-volatile ROM segment.
} >ram AT >rom /* VMA in ram, LMA in rom */


Now that we have our linker configured to locate our initialization code in the right place, and are exporting the addresses of important features such as the global pointer and initial stack pointer, we can start writing the lowest level code for our system. This startup code will initialize our stack, so that we can safely make function calls, the global pointer, so that linker-relaxed addressing works, and then call into our next section of initialization code:

# Since we need to ensure that this is the very first code the CPU runs at
# startup, we place it in a special reset vector section that we link before
# anything else in the .text region
.section .reset_vector

# In order to initialize the stack pointer, we need to know where in memory
# the stack begins. Our linker script will provide this symbol.
.global _stack

# Our main application entrypoint label

# Initialize global pointer
# Need to set norelax here, otherwise the optimizer will convert this to
# mv gp, gp which wouldn't be very useful.
.option push
.option norelax
la gp, __global_pointer$
.option pop

# Load the address of the _stack label into the stack pointer
la sp, _stack

# Once the register file is initialized and the stack pointer is set, we can
# jump to our actual program entry point
call reset_handler

# If our reset handler ever returns, just keep the CPU in an infinite loop.
j loop

Once the assembly code above has initialized the state of the gp and sp registers, we are able to safely start calling methods and executing code beyond careful assembly, so for the next few initialization steps we jump to the reset_handler method, which we will write in C:

// In our linker script, we defined these symbols at the start of the
// region in ROM where we need to copy initialized data from, and at the start
// and end of the data section that we need to copy that data to
extern unsigned _data_loadaddr, _data, _edata;

// Likewise, the locations of the preinit, init and fini arrays are generated
// by the linker, so we need to tell the compiler that they are defined
typedef void (*void_fun)(void);
extern void_fun __preinit_array_start, __preinit_array_end;
extern void_fun __init_array_start, __init_array_end;
extern void_fun __fini_array_start, __fini_array_end;

// In order for the reset_handler symbol to be usable by the assembly above, we
// need to protect it from the C++ name mangler.
// Likewise, we need to assert that there exists somewhere an application
// main() that we can invoke.
extern "C" {
int main(void);
void reset_handler(void);

void reset_handler(void) {
  // Load the initialized .data section into place
  volatile unsigned *src, *dest;
  for (src = &_data_loadaddr, dest = &_data; dest < &_edata; src++, dest++) {
    *dest = *src;

  // Handle C++ constructors / anything with __attribute__(constructor)
  // These regions contain an array of function pointers, so we simply need to
  // iterate each and invoke them
  void_fun *fp;
  for (fp = &__preinit_array_start; fp < &__preinit_array_end; fp++) {
  for (fp = &__init_array_start; fp < &__init_array_end; fp++) {

  // At last, we can jump to our actual application level code

  // Should our application code ever exit (unusual in embedded), we may as
  // well run the desctructors properly
  for (fp = &__fini_array_start; fp < &__fini_array_end; fp++) {

Memory-mapped IO

After building our way up from the bottom, we have finally arrived at the application level. From here, we can implement int main() and proceed to write firmware with impunity. But we didn’t come all this way just to run code in isolation, we want our code to be able to reach out and interact with the peripherals we build into our FPGA. So to that end, let’s take a look at how memory mapped IO works from the firmware side.

We saw earlier that our CPU has two buses - an instruction bus and a data bus. When it needs to fetch instructions, the instruction bus reaches out over Wishbone and makes a read transaction at the relevant address. Similarly, if a load or store is executed, the data bus will generate a Wishbone transaction against the given location. But there is nothing that states that location has to be a memory - we are free to have those reads and writes be routed to any wishbone peripheral we desire. So let’s create a simple RGB LED controller that’s accessible over wishbone, and demonstrate how we can control and query it from our C code.

The core logic of our peripheral will look very similar to the memory we implemented a little while earlier, except that instead of a large block ram we will create some number of small registers to hold the data the peripheral needs to know. In this case, let’s say that our peripheral will be an 8-bit PWM generator with red, green and blue output channels. Since our registers can be up to 32 bits wide, we can pack the RGB component into one register and use a second to control the prescaler for our PWM generation. In verilog, this might look a little like this:

// PWM prescaler register
reg [31:0] pwm_prescaler;

// BGR output compare registers
reg [7:0] ocr_b;
reg [7:0] ocr_g;
reg [7:0] ocr_r;

// Wishbone register addresses
    wb_r_PWM_PRESCALER  = 1'b0,
    wb_r_BGR_DATA       = 1'b1,
    wb_r_MAX            = 1'b1;

// Since the incoming wishbone address from the CPU increments by 4 bytes, we
// need to right shift it by 2 to get our actual register index
localparam reg_sel_bits = $clog2(wb_r_MAX + 1);
wire [reg_sel_bits-1:0] register_index = i_wb_adr[reg_sel_bits+1:2];

always @(posedge i_clk) begin
    if (i_reset) begin
        o_wb_ack <= 0;
        pwm_prescaler <= 0;
    end else begin
        // As in our RAM before, we can default our ack strobe low
        o_wb_ack <= 1'b0;

        // If we are addressed by the cyc and stb lines, and are not in the ack
        // out cycle, we are in a transaction
        if (i_wb_cyc && i_wb_stb && !o_wb_ack) begin
            // Once again we do not have any delays, so we can ack
            // unconditionally
            o_wb_ack <= 1'b1;

            // Handle register reads
            // Note that our BGR data is 24 bits, so we pad it to 32
            case (register_index)
                wb_r_PWM_PRESCALER: o_wb_dat <= pwm_prescaler;
                wb_r_BGR_DATA:      o_wb_dat <= {8'd0, ocr_b, ocr_g, ocr_r};

            // Handle register writes if the write enable flag is set
            if (i_wb_we) begin
                case (register_index)
                    wb_r_PWM_PRESCALER: pwm_prescaler <= i_wb_dat;
                    wb_r_BGR_DATA: begin
                        // For RGB writes, break out the data to the individual
                        // output compare registers
                        ocr_b <= i_wb_dat[23:16];
                        ocr_g <= i_wb_dat[15:8];
                        ocr_r <= i_wb_dat[7:0];

The full implementation of this module can be found here. Once we have granted this peripheral an entry in our wb intercon YAML and connected it up in our top module, we can turn back to the firmware and take a look at what’s necessary to interact with it. If we located our LED peripheral at wishbone address 0x40002000, then in our C++ code if we read or write the 32-bit memory at that location, what we will actually be doing is reading or writing the prescaler register of our LED block. It’s that easy! The only two things we need to make sure we are clear about in our code is that

To make following these rules simple, it helps to define some preprocessor macros as follow:

// Redefines an integer constant as a dereferenced pointer to a volatile 32-bit
// memory mapped IO register
#define MMIO32(ADDR) (*(volatile uint32_t *)(ADDR))

// Defines a register with 32-bit offset OFFSET
#define REG32(BASE, OFFSET) MMIO32(BASE + (OFFSET << 2))

// Now that we have our two macros above, we can quickly make some readable
// definitions for our LED driver
#define LED_BASE            0x40002000
#define LED_BGR_DATA        REG32(LED_BASE, 1)

With only that, we can now read and write these registers from our code like another variable. Let’s finally implement a main() and cycle some colours on our LED:

#define CPU_CLK_HZ 42'000'000

uint32_t led_states[] = {

void delay(uint32_t cycles ) [
    while (cycles--){
        asm volatile ("nop");

int main(void) {
    // Initialize our PWM prescaler to generate a 1kHz carrier with 8 bit pwm
    LED_PWM_PRESCALER = (CPU_CLK_HZ / 256 / 1'000) - 1;

    // Iterate through each of our LED states, with a delay so that we can
    // actually see what is going on
    for (int i = 0; i < sizeof(led_states) / sizeof(uint32_t); i++) {
        LED_BGR_DATA = led_states[i];
        delay(CPU_CLK_HZ / 2); // Very approximately one second


While our above demo is very blinky, it’s not particularly elegant - our delay loop just burns CPU cycles, and it would be hard to do cycle counting if we were trying to achieve a number of timed tasks at the same time. Instead it would be a lot more useful if we could somehow keep track of the current time, and only advance the LED state after a given number of milliseconds.

Luckily, building a timer is very much within our reach - with enough gates we can build anything! But how do we tie the timer back into our CPU without just busy waiting on a timer register instead of a NOP loop? One solution to this problem is interrupts. If we build a timer module and connect it to one of the interrupt lines on our CPU, our CPU can then jump to an interrupt handler that either updates our LEDs directly, or simply updates a counter we can reference from our main loop.

Let’s start with the verilog timer module. We don’t need anything too fancy, but we do want to be able to at least configure the prescaler for the counter on the fly, so that we can adjust the interrupt rate in our firmware. We also need to be able to clear the interrupt signal from the timer, so it’s back to the Wishbone peripheral structure we are getting increasingly familiar with.

The first part of a timer is simple: we need to be able to count. In this case, we count down from our configured prescaler to zero, then reload the prescaler and start all over again, something like this:

// Prescaler value. Reloaded onto the downcounter on update.
reg [31:0] prescaler = 32'hFFFF_FFFF;;

// Downcounter. Trigger output is latched high when this hits zero.
reg [31:0] downcounter = 32'hFFFF_FFFF;;

// Trigger output signal
reg timer_trigger = 0;

// Downcount until the counter reaches zero, then reload the prescaler and
// start counting down again
always @(posedge i_clk) begin
    if (downcounter > 0) begin
        downcounter <= downcounter - 1;
    end else begin
        downcounter <= prescaler;
        timer_trigger <= 1'b1;

Now that we have that, we need to add enough wishbone logic to allow for setting the prescaler, and for clearing the timer trigger when it is set. Since interrupts on the CPU are level-triggered, if we don’t clear the interrupt source, the CPU will end up returning from and re-entering the timer interrupt forever!

// Wishbone register addresses
// Each register is 32 bits wide
    wb_r_PRESCALER  = 1'b0,
    wb_r_FLAGS      = 1'b1,
    wb_r_MAX        = 1'b1;

// Bit indices for the flags register
    wb_r_FLAGS__TRIGGER = 0;

// Since the incoming wishbone address from the CPU increments by 4 bytes, we
// need to right shift it by 2 to get our actual register index
localparam reg_sel_bits = $clog2(wb_r_MAX + 1);
wire [reg_sel_bits-1:0] register_index = i_wb_adr[reg_sel_bits+1:2];

always @(posedge i_clk) begin
    // Default the ack signal to a zero state.
    // Later writes to this register will take precedence if we are actually
    // performing a wishbone transaction
    o_wb_ack <= 1'b0;

    // If the cycle and strobe inputs are high, and this is not the cycle after
    // a previous transaction, we are servicing an actual wishbone request
    if (i_wb_cyc && i_wb_stb && !o_wb_ack) begin
        // None of our operations take more than one cycle, so we can always
        // unconditionally ack the request
        o_wb_ack <= 1'b1;

        // To handle writing the prescaler / clearing flags, we need to check
        // if this is a write request
        if (i_wb_we) begin
            // If it is, use the address bits to select the appropriate
            // register to work with
            case (register_index)
                wb_r_PRESCALER: begin
                    // Load the new prescaler, and also reset the downcounter
                    // If we don't reset the downcounter, we run the risk that
                    // the previously loaded value was extremely large and will
                    // delay 'proper' operation at the new prescaler rate
                    prescaler <= i_wb_dat;
                    downcounter <= i_wb_dat;
                wb_r_FLAGS: begin
                    // If this is a write to the flags register, we want to
                    // check which flags are being cleared.
                    // If the trigger bit is written, we clear the trig state.
                    if (i_wb_dat[wb_r_FLAGS__TRIGGER])
                        timer_trigger <= 1'b0;

Note - in the two preceding snippets, the timer_trigger register is written to from two different always blocks - this is not valid verilog! If you are going to copy parts of this module, please do so from the full source code.

Now that we have a timer implementation, the only remaining change to make in gateware is connect up the timer trigger output wire to the timer interrupt input on our CPU, and to connect the timer to the wishbone bus wires we generated with our wb_intercon file.

// Timer for generating the timer interrupt
wire timer_interrupt;
timer_wb #(
) timer0 (

// CPU
VexRiscv cpu0 (
    // Interrupt sources

With that hooked up, our gateware is good to go! Let’s take a quick look at our simulation and see that interrupt in action:

Timer interrupt execution?

Except… we see nothing! Our timer counts down and sets the trigger output, but our CPU ticks along as though nothing has happened. What gives! The clue here is the additional signals on screen - the default value of the Configuration and Status Registers (CSRs) that relate to interrupts have all interrupts disabled by default. So even though we see that the Machine Timer Interrupt Pending (MTIP) flag is set, nothing happens. So we need to switch over to our firmware and make sure that we configure the CPU properly to handle our interrupts.

The first register we need to configure is the Machine Trap-Vector Base-Address register, or mtvec. This register is used to store the memory address that the CPU will jump to in the event of an interrupt. Since the address must be 32-bit aligned, the low 2 bits are used to control the interrupt mode, where a value of 0x0 is direct (all exceptions jump to the mtvec address) and a value of 0x1 is vectored (interrupts set the program counter to (mtvec + (4 * exception_code)). In vector mode, mtvec must therefore be the start of a series of jump statements to specific exception handlers, instead of the entry point to a single exception handler. For simplicity we will use the non-vectored version and work out how to handle the interrupt in software. But before we can set the mtvec address at all, we need some sort of interrupt handler routine to point it to, so let’s create one now.

// We need to decorate this function with __attribute__((interrupt)) so that
// the compiler knows to save/restore all register state, as well as to
// re-enable interrupts on return with the mret instruction.
void __attribute__((interrupt)) interrupt_handler(void) {
    // When an interrupt occurs, the mcause register contains the interrupt type
    uint32_t mcause;
    asm volatile("csrr, %0" : "=r"(mcause));

    // The top bit of mcause is the sync vs async exception bit, we don't
    // handle that here so mask it off
    mcause &= 0x7FFFFFFF;

    // If the cause is some number out of range of our handler table, we have
    // no way to handle this interrupt! Block forever.
    if (mcause >= (sizeof(vector_table) / sizeof(isr_vector))) {
        while (true) {}

    // Otherwise, we can jump to the handler listed in our vector table.
    // Since we took care to order our struct to match the interrupt IDs, we can
    // reinterpret it as an array for easy indexing based on mcause
    ((isr_vector *)&vector_table)[mcause]();

Now you’ll note that in the code above we made some reference to a vector table - since we are decoding the interrupt cause ourselves, we need some structure that contains pointers to the handlers for each of our interrupts. To keep things well named, we use a struct that has the same layout as the interrupt numbering scheme used by RISC-V, like so:

// Create a type alias for our exception handlers, which are void functions
typedef void (*isr_vector)(void);

// The basic interrupts for RISC-V are the software, timer and external
// interrupts, each of which is specified for the user, supervisor and machine
// privilege levels.
struct {
    // Software interrupt
    isr_vector software_user_isr = &blocking_handler,
    isr_vector software_supervisor_isr = &blocking_handler,
    isr_vector software__reserved = &blocking_handler,
    isr_vector software_machine_isr = &blocking_handler,
    // Timer interrupt
    isr_vector timer_user_isr = &blocking_handler,
    isr_vector timer_supervisor_isr = &blocking_handler,
    isr_vector timer__reserved = &blocking_handler,
    isr_vector timer_machine_isr = &timer_interrupt,
    // External interrupt
    isr_vector external_user_isr = &blocking_handler,
    isr_vector external_supervisor_isr = &blocking_handler,
    isr_vector external__reserved = &blocking_handler,
    isr_vector external_machine_isr = &blocking_handler,
} vector_table;

This struct, when compiled, turns into an array of function pointers that we can index directly with mcause. For interrupts that we don’t explicitly handle yet, we point to a simple blocking handler that loops forever, hanging the CPU so that we can detect something has gone wrong. The one interrupt that we do actually want to handle properly right now, the timer interrupt, is pointer to a timer_interrupt function, which we will define like so:

// Our counter is modified from an interrupt context and read from a
// non-iterrupt context, so we need to mark it volatile to ensure accesses hit
// main memory each time
volatile uint32_t time_ms = 0;
void timer_interrupt(void) {
    // Increment the millisecond counter

    // We need to also clear the source of the interrupt, otherwise when we
    // return from interrupt it will just fire again right away.

Alright, we now have all the handler code that we should need once our interrupt actually fires. However, we still have yet to actually enable interrupts or tell the CPU where exactly to go when an interrupt is triggered. In order to do this, let’s add some code early in our firmware initialization that sets the relevant CSRs:

// Firstly, let's update the machine interrupt vector CSR with the address of
// our base interrupt handler.
// Since our address is 32-bit aligned, we are in non-vectored mode by default
asm volatile("csrw mtvec, %0" ::"r"(&interrupt_handler));

// We now want to enable the machine timer interrupt. To do this, we need to
// set bit 7 in the Machine Interrupt Enable (mie) register
// We use the assembler shorthand CSR Set bit instruction here for convenience
// Note that the literal 1<<7 is here marked as a register operand - the CSRS
// instruction only supports 5-bit immediates
asm volatile("csrs mie, %0" ::"r"(1<<7));

// We then need to enable machine interrupts globally, by setting bit 3 in the
// Machine Status Register (mstatus).
// Here, 1<<3 fits in the immediate form of the instruction ("i")
asm volatile("csrs mie, %0" ::"i"(1<<3));

Finally, we should have completed all the necessary legwork to get timer interrupts running on our CPU. To verify, let’s take another look at our simulated system:

Timer interrupt execution with MTIE and MIE enabled

Much better! We can see that the MTIE and MIE bits are now set, and that we have a clear cycle by cycle progression of our timer output signal going high, being registered in the CPU as a pending timer interrupt (MTIP), the corresponding interrupt code (7) being loaded in mcause, and finally the CSR-triggered jump to our interrupt handler at 0x200000190. If we were to zoom out on the scope above, we would later see the timer trigger be cleared by the write in our timer interrupt handler to the timer flags register, the deassertion of MTIP and the return to normal code execution.

If we build and deploy this with a couple of tweaks to our colour generation (see full code here) we should end up with a lovely changing colour cycle running on our dev board:

Interrupt-powered RGB cycling

The full source code for all gateware, firmware, tests and simulations used in this post can be found on Github here. The repo also contains some code and peripherals that we did not cover here, such as a buffered UART peripheral.

Design files and assembly instructions for the FPGA board shown in the image above can be found in this previous post.

FPGAs 2: Multiplexed RGB LED Panels

Companion code for this post available on Github

One of the more eyecatching demonstrations of FPGAs recently has been the trend to show off smoothly animating LED Cubes of various sizes and shapes. But what goes into making these? How do we interface with the LED panels that construct them? And can we use those panels to show an animated GIF? In this post, I will attempt to cover the building blocks necessary to load custom animations onto a FPGA board powered by the Lattice HX4K, and display those animations on a 64x64 LED panel. All source code for the following can be found on Github.

How do RGB panels work?

While the specific layout of the various LED panels available online may differ, the general architecture is by and large the same. Three sets of shift registers (one each for the red, green and blue pixel data) have each bit of their output connected across one column of the LED grid. This means that whatever bit pattern is shifted into these registers controls which columns of the display will light up. However, we generally don’t want to the same data on every line - in order to select a single row for the given column data, the common pin of each LED in a row is connected to an output of an address demultiplexer, which uses the address select inputs to enable only a single row at a time. A simplified diagram of a 64-column by 32-row display is shown below (many rows/cols have been omitted):

Subsection of common RGB panel

One thing to note about this architecture is that while it is simple in terms of component cost, we can only actually display one unique row of data at a time. In order to show an entire image, we need to quickly scan through displaying rows at some speed above which the strobing is invisible to the naked eye, and the persistence of vision effect makes the image appear solid. This sort of high speed data shifting application is a perfect place for FPGAs, so in the rest of this post we will try and cover getting from displaying a solid colour to full animated GIFs with 5 bit colour channels.

If you want to follow along, I would strongly suggest building one of the ICE40 HX4K Breakout Boards I wrote about here, which the code samples in this post have been written for, or a comparable board such as 1bitsquared’s iCEBreaker.

You will also need a LED panel - the one I will be using here is a 64x64 model with a HUB75E connector. You can find similar panels cheaply on AliExpress (search HUB75E, 64x64), or purchase find them at various other sellers on the internet. This panel is internally organized as two 64x32 panels, so instead of having just the one set of RGB shift registers it instead has two, one for the top half and one for the bottom. This doesn’t mean too much for our design, we will just have two sets of outputs.

Finally, you will need some way to connect your FPGA board to the display. You can either use jumper wires or a PMOD style breakout, such as this one.

Getting started

The first thing to do with any project is get to the point where you can see everything working (or not working) end to end. So let’s start by creating a panel driver module that will cycle through displaying each of the 8 primary colours (the possible permutations of R, G, B) in turn to make sure that we are able to shift data in and advance the row selector. The first step in this is to configure which pins on our FPGA board connect to what signals of our LED panel. For me, I have the PMOD to HUB75E board connected to PMOD ports C1 and C2 on the HX4k Breakout board, as shown:

PMOD to HUB75E adapter

This means that I can use the PMOD[] definitions from the Placement Constraint File for the board when interfacing with the RGB panel. So in my top level verilog file, I can declare my inputs like so:

module top(
        // Input clock
        input wire CLK_12MHZ,

        // PMODs
        output wire [7:0] PMOD_C1,
        output wire [7:0] PMOD_C2

    // Map those PMOD pins to some real signal names
    wire [1:0] rgb_panel_r = {PMOD_C1[4], PMOD_C1[0]};
    wire [1:0] rgb_panel_g = {PMOD_C1[5], PMOD_C1[1]};
    wire [1:0] rgb_panel_b = {PMOD_C1[6], PMOD_C1[2]};
    wire [4:0] rgb_panel_a = {PMOD_C2[7],
    wire rgb_panel_bl = PMOD_C2[4];
    wire rgb_panel_la = PMOD_C2[5];
    wire rgb_panel_ck = PMOD_C2[6];

With our IO named in the top module, let’s create a second module to hold all of the logic needed to drive out display, called panel_driver. To start, we will just try and display solid colours, one after the other. Here is the full verilog description of that module:

`default_nettype none
`define CLOCK_HZ 12_000_000
module single_colour_test(
    // Input clock to our panel driver
    input wire i_clk,
    // Shift register controls for the column data
    output reg o_data_clock,
    output reg o_data_latch,
    output reg o_data_blank,
    // Data lines to be shifted
    output reg [1:0] o_data_r,
    output reg [1:0] o_data_g,
    output reg [1:0] o_data_b,
    // Inputs to the row select demux
    output reg [4:0] o_row_select

    // How many pixels to shift per row
    localparam pixels_per_row = 64;

    // State machine IDs
        s_data_shift = 0,
        s_blank_set = 1,
        s_latch_set = 2,
        s_increment_row = 3,
        s_latch_clear = 4,
        s_blank_clear = 5;

    // Simple colour cycling logic. We will have a prescaler that counts down
    // to zero twice per second, based on the frequency of our module input
    // clock (`CLOCK_HZ`).
    // Whenever this countdown hits zero, we will increment our colour state
    // register, each bit of which is mapped to the reg, green or blue data
    // channel of the RGB panel shift registers.
    localparam COLOUR_CYCLE_PRESCALER = (`CLOCK_HZ / 2) - 1;
    reg [$clog2(COLOUR_CYCLE_PRESCALER):0] colour_cycle_counter = 0;
    reg [2:0] colour_register;
    always @(posedge i_clk) begin
        if (colour_cycle_counter == 0) begin
            colour_register <= colour_register + 1;
            colour_cycle_counter <= COLOUR_CYCLE_PRESCALER;
        end else
            colour_cycle_counter <= colour_cycle_counter - 1;

    // Connect the output colour data lines to our colour counter
    assign o_data_r = {colour_register[0], colour_register[0]};
    assign o_data_g = {colour_register[1], colour_register[1]};
    assign o_data_b = {colour_register[2], colour_register[2]};

    // Register to keep track of where we are in our panel update state machine
    reg [2:0] state = s_data_shift;
    // How many pixels remain to be shifted in the 'data_shift' state
    reg [7:0] pixels_to_shift = 64;
    always @(posedge i_clk) begin
        case (state)
        s_data_shift: begin // Shift out new column data for this row
            if (pixels_to_shift > 0) begin
                // We have data to shift still
                if (o_data_clock == 0) begin
                    // For this test, we have hardcoded our colour output, so
                    // there is nothing to do per-pixel here
                    o_data_clock <= 1;
                end else begin
                    o_data_clock <= 0;
                    pixels_to_shift <= pixels_to_shift - 1;
            end else
               state <= s_blank_set;
         // In order to update the column data, these shift registers actually
         // seem to require the output is disabled before they will latch new
         // data. So to perform an update, we have a series of steps here that
         // - Blank the output
         // - Latch the new data
         // - Increment to the new row address
         // - Reset the latch state
         // - Unblank the display.
         // Each step has been made it's own state for clarity; if one wanted
         // to save a little more on logic some of these steps can be merged.
         s_blank_set: begin o_data_blank <= 1; state <= s_latch_set; end
         s_latch_set: begin o_data_latch <= 1; state <= s_increment_row; end
         s_increment_row: begin o_row_select <= o_row_select + 1;
                                state <= s_latch_clear; end
         s_latch_clear: begin o_data_latch <= 0; state <= s_blank_clear; end
         s_blank_clear: begin
             o_data_blank <= 0;
             pixels_to_shift <= pixels_per_row;
             state <= s_data_shift;

If we build and deploy this to our FPGA test board, we should see the connected LED panel flash through the 8 possible colours we have with each color channel being fully on / off:

Full-colour test pattern

Displaying Images

So far so good - we are now able to control the colours output to the panel, and can do so fast enough that the panel appears to be solid. However, so far we are just displaying solid colours - not very useful. Instead, let’s add the ability for our gateware to display one of the most popular image formats: the animated GIF.

Our test image: nyancat

But this begs the question, how do we get all that data loaded into our FPGA? And once it’s in there, how do we display it? If we were to store each frame of the above GIF as a 64x64 array of bytes, we would need 64 * 63 * 3 * 11 frames of data - that’s 132 KiB. Looking at the datasheet for our HX4K, we see that it only has 80K ram bits - only 10K ram bytes! We will clearly need to be more creative. If instead of using a full byte per colour pixel we shave each channel down to 5, we can store one pixel in two bytes, using 5 bits each for Red/Green/Blue -> 15 bits per pixel. This brings us down to 88KiB for the entire animation, still far too large. However, it does bring us down to 8KiB per frame - which does fit in our 10KiB of system memory. But that still only fits one frame - in order to display an animation, we will have to have some persistent storage that we can read each frame from into our memory buffer so that we can blast it to the screen. Luckily, on this board, we have such a persistent storage - the flash chip used for storing the FPGA bitstream.

The HX4K bitstream is, at maximum, 136,448 bytes (according to Lattice note TN1248) . However, the flash chip on the HX4K breakout we are using here is 128 megabits, or 16 megabytes. This means that even with the bitstream using the lower part of the flash, we can store image data in the higher address regions so that our application can read out frame data when it needs it. So let’s write a quick program that, given a series of input images that are 64x64 pixels, converts each one to a packed RGB565 and then writes all of the frames, in order, to a binary file that we can then load onto the flash of the FPGA board. The full source of this utility, img_pack, is included in the Github repo that accompanies this post. In order to use it on our test image, we first need to do a little preprocessing using ImageMagick.

# Resize the gif to 64x64 pixels, ignoring aspect ratio
convert -resize 64x64\! nyan_fullres.gif nyan_64.gif
# Convert the resized gif into individual frames
# This will generate nyan_frame-0.png through nyan_frame-11.png
convert -coalesce nyan_64.gif nyan_frame.png
# Now that we have our individual frames as images, we can use the img_pack
# tool to generate a binary representation of the gif to be loaded to our board
img_pack nyancat.bin nyan_frame-{0..11}.png

After that, if you are using the HX4K dev board you can use the faff tool to load the program into flash at a safe offset that won’t interfere with the FPGA bitstream itself. If you are using the iCEBreaker board, the iceprog has a similar option for loading data at a specific offset.

# To load the data into flash, we can use the faff utility to interact with the
# HX4K dev board, and write the data at offset 0x800000, safely out of the way
# of the bitstream
faff --lma 0x800000 -f nyancat.bin --no-verify
# Alternately, on an iCEBreaker, the following should work:
iceprog -o 0x800000 -X nyancat.bin

Reading from Flash

Now that the data is safely on the flash, we need to be able to actually access it from our FPGA, and then store it in our RAM buffer so that our panel driver can then read it and blit it to our display. The flash chip on the HX4K board, as well as those on almost all other flash chips, share a common set of command words. All of them can be found in the datasheet (see here), but the important ones here are:

  1. 0x66, 0x99: Reset enable / reset. This command series will put the device back into a sane state to receive commands, important to do at boot since we can’t gauarantee what operation the flash was in the middle of whenever our gateware came online.
  2. 0xAB: Release from power down. The HX4K, after loading the bitstream, is careful to put the flash chip into a power down state, since as far as it’s concerned it’s done with the flash until it next needs to load the bitstream. In order to read our frame data from it, we need to send this instruction to first wake it up.
  3. 0x0B: Fast read. To get our pixel data out, we send this command followed by the 24 bit start address for our read, and then after one byte worth of delay the flash chip will output it’s data, one byte at a time, until we stop reading. When we need to read the next frame, we just send another fast read command and a new address, and we are good to go.

With these commands, we can think of our flash loader state machine as being:

  1. Idle state. Wait here until we get a signal to load the next frame from flash.
  2. Reset state. Select the flash chip, clock out the reset command.
  3. Wakeup state. Select the flash chip, and release it from power down.
  4. Initiate read state. Select the flash chip, and clock out the fast read command followed by the address data and the idle byte.
  5. Perform read. For as many bytes as are in a frame, shift in data and store it in our ram buffer. When we are done, return to the idle state.

A shortened version of the flash loader verilog is shown below. The full code listing can be found here.

// Flash commands for use later
`define FLASH_OP_RESET    16'H6699
`define FLASH_OP_WAKEUP    8'HAB

always @(posedge i_clk) begin
    case (state)
        // Default idle state. This state will only transition if the 'initiate
        // read' module input is set, at which point it will latch the base
        // address and start the read operation
        s_idle: begin
            if (i_read_stb) begin
                flash_read_addr <= i_read_addr;
                state <= s_initiate_reset;

        // Perform a flash chip reset. Takes advantage of our
        // flush_command_buffer state, by just specifying the data to shift,
        // and which state to transition to after the command operation is
        // complete.
        s_initiate_reset: begin
            // Our command buffer just needs the two reset bytes, the rest
            // of the bits are don't care (zero)
            command_buffer <= {`FLASH_OP_RESET, 24'b0};
            cmd_buffer_bits_to_shift <= 16;
            flash_cs <= 0;
            // Once the command is done, we want to move to the wakeup step
            next_state <= s_initiate_wakeup;
            // Directly transition to the generalized 'flush buffer' state.
            state <= s_flush_command_buffer;

        // Similar to the above, but takes the flash chip out of sleep mode
        s_initiate_wakeup: begin
            command_buffer <= {`FLASH_OP_WAKEUP, 32'b0};
            // [...]
            state <= s_flush_command_buffer;

        // Once again, we set up the data for flush_command_buffer, this time
        // with the 'fast read' command and the start address we latched when
        // we started the read process
        s_initiate_read: begin
            command_buffer <= {`FLASH_OP_FAST_READ, flash_read_addr, 8'b0};
            // [...]
            state <= s_flush_command_buffer;

        // State for performing a command action. Shifts out a given number of
        // bits from the command buffer (MSB first), and when done transitions
        // to the state idenfitied by the next_state register.
        s_flush_command_buffer: begin
            if (cmd_buffer_bits_to_shift > 0) begin
                if (flash_sck == 0) begin
                    // Latch new data and perform rising edge
                    {flash_mosi, command_buffer} <= {command_buffer, 1'b0};
                    flash_sck <= 1;
                end else begin
                    // Perform falling edge and decrement bit count
                    flash_sck <= 0;
                    cmd_buffer_bits_to_shift <= cmd_buffer_bits_to_shift - 1;
            end else begin
                // Done - move to continuation state
                state <= next_state;
                // If we aren't moving to the read state, deassert CS
                if (next_state != s_shift_data)
                    flash_cs <= 1;

        // Here is where we actually handle the incoming data from the flash
        // chip. Each 16 bit word is one pixel, so every time we successfully
        // read 16 bits we write it out to a block memory using the
        // ram_address, ram_data and ram_write_enable signals.
        s_shift_data: begin
            // We are going to care about 16 bit chunks, since that's
            // the interface we're using for our block RAMs
            if (words_to_read == 0) begin
                ram_write_enable <= 0;
                flash_cs <= 1;
                state <= s_idle;
            end else begin
                if (input_bits_to_shift == 0) begin
                    // Done shifting a word, move it to the output data
                    // lines and strobe the write signal
                    ram_data <= input_shift_data;
                    ram_write_enable <= 1;
                    words_to_read <= words_to_read - 1;
                    input_bits_to_shift <= 16;
                end else begin
                    // If we just did a write, bring the strobe back down
                    // and increment the write address for next time
                    if (ram_write_enable) begin
                        ram_write_enable <= 0;
                        ram_address <= ram_address + 1;
                    if (o_flash_sck == 0) begin
                        // Set up rising edge of SPI clock
                        flash_mosi <= 1'b0;
                        flash_sck <= 1;
                    end else begin
                        // Falling edge
                        flash_sck <= 0;
                        input_shift_data <= {input_shift_data[14:0], i_flash_miso};
                        input_bits_to_shift <= input_bits_to_shift - 1;

Simulating our Flash reader

Since reading from a flash chip is slightly more involved than shifting out our pixel data, and a little harder to visually debug, let’s write a quick test interface in Verilator so that we can observe that our flash interface sends the right data with the right timings to initialize a flash read, and that it then takes the incoming data and stores it in the right place. Full source code for the testbench can be found here.

To make our I/O more realistic, in our testbench we can create a fake SPI chip that responds to read requests by keeping track of the output signals from the flash_loader verilog module. An annotated excerpt of the flash simulator is shown below:

// Data to to 'read'
const uint8_t *data =
    reinterpret_cast<const uint8_t *>("This is a test string");
const unsigned data_len = strlen(data);

// Fake flash update method, should be called after each master clock cycle of
// the simulation
void DummyFlash::eval() {
  // Did the CS state change?
  if (*spi_cs != last_cs) {
    fprintf(stderr, "CS state: was %u, now %u\n", last_cs, *spi_cs);
    // If the cs line went low on this clock cycle, then the flash chip is now
    // selected. Use this as a trigger to reset our fake flash state.
    if (*spi_cs == 0) {
      shift_byte_index = 0;
      shift_bit_index = 0;
      *spi_miso = 1;
      last_sck = *spi_sck;
    } else {
      // Deselected. Move the miso line to a known state.
      *spi_miso = 1;
    // Keep track of changes to the chip select pin.
    last_cs = *spi_cs;

  // If we aren't selected, just return
  if (last_cs == 1) {

  // Did the clock transition?
  if (*spi_sck != last_sck) {
    last_sck = *spi_sck;
    if (last_sck) {
      // Posedge, increment the number of bits we have received from the
      // SPI master
    } else {
      // Negedge
      // Don't shift data if this is the read preamble
      if (preamble_bit_count >= 40) {
        // Set miso to be the next bit
        *spi_miso =
            data[shift_byte_index % data_len] &
              (0b1000'0000 >> shift_bit_index) ? 1 : 0;
        // Increment the bit index
        // If we hit the end of the byte, increment the byte index
        if (shift_bit_index > 7) {
          shift_bit_index = 0;
          fprintf(stderr, "New byte: %02x\n", data[shift_byte_index]);

In this simple flash example, we don’t actually verify the input address / reset state of the flash, but hopefully this example gives an idea of how the test code might be extended to perform this checking. If we then compile and run the full testbench with tracing enabled, like so:

# Build the CMake project that contains the Verilator testbenches
mkdir tb/build
cd tb/build
cmake ../
make -j$(nproc)
# Run the flash loader testbench. This will generate the flash_loader.vcd file,
# which can be examined in GTKwave

We get an output file that can be loaded into GTKWave, where we can observe the communication between our flash loader verilog, our emulated flash and the output signals to our pixel buffer memory. By looking carefully at these signals, we can be confident that our design will work once deployed on our actual FPGA board.

Trace of simulated flash loader reading from 'Flash'

Rendering from memory

With a module for reading our pixel data from flash, the only things remaining to get animated images running on our LED panel is an animating counter for triggering the load of the next frame, and then a modification to the panel driver code to use the loaded frame data.

// Animation counter for triggering frame loads from the flash chip
localparam ANIMATION_UPDATE_PRESCALER = (`clk_hz / 10) - 1; // 10fps
localparam ANIMATION_FRAME_COUNT = 11; // Number of frames in the animation
reg [$clog2(ANIMATION_UPDATE_PRESCALER):0] prescaler_reg = 0;
always @(posedge i_clk) begin
    if (prescaler_reg == 0) begin
        // Reload the prescaler downcounter register
        // Strobe the flash load signal
        flash_load_strobe <= 1;
        // If we have reached the final frame in the animation, go back to
        // frame 0. Otherwise, just advance to the next frame.
        if (frame_index == ANIMATION_FRAME_COUNT)
            frame_index <= 0;
            frame_index <= frame_index + 1;
    end else begin
        // Downcount prescaler
        prescaler_reg <= prescaler_reg - 1;
        flash_load_strobe <= 0;


// To connect our frame counter up to the flash reader, and knowing that each
// frame is 64*64*2 = 8192 bytes, we can create a wire that counts up the start
// address of each animation frame from a specified base address.
localparam FLASH_BASE = 24'h80_00_00;
wire [23:0] flash_load_addr = {FLASH_BASE[23:21], frame_index, 13'b0};

For our panel driver, we can change our data output from our solid colour registers we used earlier to the most significant red, green and blue bit of the pixel byte. This will be enough for us to view on the panel whether we have succeeded in loading the frame data from flash.

// On the negative edge of the serial clock cycle, set the output
// red/green/blue data. For this intermediate step, we will just set the pixel
// value to the most significant bit of that channel.
if (data_clock == 0) begin
    // We need to load the n'th most significant bit of each colour channel,
    // based on which pixel bit index we are currently displaying
    // pixel_bit_index is in range 0..4
    data_r <= {i_ram_b2_data[11 + pixel_bit_index],
               i_ram_b1_data[11 + pixel_bit_index]};
    data_g <= {i_ram_b2_data[6 + pixel_bit_index],
               i_ram_b1_data[6 + pixel_bit_index]};
    data_b <= {i_ram_b2_data[0 + pixel_bit_index],
               i_ram_b1_data[0 + pixel_bit_index]};
    data_clock <= 1;
    ram_addr <= ram_addr + 1;

With this updated gateware, if we rebuild and reflash our FPGA board we should be greeted by something vaguely recognizable as our desired GIF output:

MSB-Only Animation running off the FPGA

Achieving colour depth

While we have managed to get an animation running on our LED panel, it’s clear that it’s not perfect yet - we are still only using our initial 8 colours, since each pixel is only displaying one bit of red green or blue colour. To improve our colour representation, and make the most of all that extra pixel information we packed into the flash, we need to get a bit smarter with our multiplexing. We want to make sure that all 5 of our data bits are used, but we also need to make sure that our multiplexing scheme reflects the relative importance of each bit. Since the colour is a binary number, each bit represents a value twice that of the less significant bit, so we should show the most significant bit of colour information for twice as long as the next most significant, and bit 1 should be shown for twice as long as bit 2, and so on and so on. This means that instead of scanning the entire display once per frame, we are actually going to scan it 5 times - the first time, we will display the most significant bit of data for 100% of the time it takes us to display the next row. Then, we will scan through with the next most significant bit, at a 50% duty cycle. The bit after that at 25%, and so on.

To achieve this, we need to add just a little more code to our panel_driver module:

// Our serial data output step takes 128 clock cycles (2 * the number of pixels
// in a row). So, if we enable the display of the row data for as long as it
// takes us to shift the new row of data, that's 100% duty cycle. If instead we
// cut off the display after 64 cycles, that's 50% duty, and so on.
// Here we'll create a short lookup table for how many cycles the display
// should be enabled, based on which bit of the colour information is being shown.
reg [8:0] time_periods_for_bit[5];
initial begin
    time_periods_for_bit[4] = 128;
    time_periods_for_bit[3] = 64;
    time_periods_for_bit[2] = 32;
    time_periods_for_bit[1] = 16;
    time_periods_for_bit[0] = 8;

// Countdown register for displaying our row data. When this hits zero, we shut
// off the output enable for the column drivers.
reg [8:0] time_periods_remaining;

// Which bit of the pixel data are we currently displaying? Counts down from
// the most significant bit (4)
reg [2:0] pixel_bit_index = 4;

// [...]

// In our data shift state, we now have one extra check to perform - decrement
// the exposure time periods, and if we hit zero, disable the output
s_data_shift: begin
    // Expose the previous row while we do our data loading.
    // Our most significant bit display time is such that it
    // matches the number of cycles it takes to load data, and then
    // each successive bit halves that value.
    if (time_periods_remaining == 0) begin
        // Turn the blanking back on
        data_blank <= 1;
    end else begin
        time_periods_remaining <= time_periods_remaining - 1;

// [...]

// We also have a little logic to add to our blank_clear state -
s_blank_clear: begin
// Load the number of time periods that we should expose this
// row for
time_periods_remaining <= time_periods_for_bit[pixel_bit_index];

// If the current row address is zero, we have done one
// full scan through the display, and should move to the
// next most significant bit of colour data
if (row_address == 0) begin
   if (pixel_bit_index == 0)
       // If we hit the lsb, wrap to the msb
       pixel_bit_index <= 4;
       pixel_bit_index <= pixel_bit_index - 1;

And with that, we should be able to render a nice, smooth animation to our LED panel.

Nyancat in glorious 5-bit colour

As always, full source code to replicate the results of this post is available on Github. If you want more details on the FPGA board used in this post, you can read about it here or take a look at the design files directly here. If you want a refresher on what exactly an FPGA is and does, this post may be useful. If you want to take things to the next level with a soft RISC-V CPU and build your own SoC, click here.

FPGAs 1: Running on Hardware

Companion code for this post available on Github

After going over the basic idea of what an FPGA is and running some simulations in the last post, the next logical step is to take things from the computer and into the real world. To that end, this post will attempt to cover the entire journey from PCB design to running gateware on an assembled board in such a way as to make it accessible to those new to FPGA development. Some implementation details have had to be left out for brevity, but every piece of CAD, firmware, and gateware referenced in this post can be found in this git repo for those that want to read deeper.

The Hardware

The hardware designed to accompany this post is a breakout board consisting of a Lattice ICE40 series HX4K FPGA, one of the few FPGAs that comes in an easy(er) to solder quad flat leaded package. The FPGA IO are broken out to 10 standard PMOD connectors, for a total of 80 external IO lines. There is also an ARM based STM32F070CBT6 microcontroller, with two hardware SPIs connected across to the FPGA (one shared with the flash), one hardware UART connected to the FPGA and a USB interface for communicating with a host computer. Full hardware design files can be found in the hardware/ folder. Feel free to use this design as a reference for further work, but if you do, bear in mind a couple of design points.

The first is that (most) FPGAs are volatile - they do not maintain their configuration if they are power cycled, and so every time they turn on they must be reconfigured before they will operate. This can be done automatically by the FPGA acting as a SPI master to read from a flash device, by having the FPGA acting as a SPI slave and receiving the bitstream data from a microcontroller or other device, or over JTAG.

To make turning on our FPGA board as simple as possible, we will include a flash chip connected to the FPGA to provide storage for the bitstream. However, we also need some way to program the bitstream onto the flash chip. One approach is to place an FTDI device such as the FT2232 on the board, which allows reading/writing the flash using MPSSE instructions. However, these chips are a bit expensive (Over $6 in quantity 1), and the functions they are performing aren’t overly complex. This is why for this design we instead use a cheap ARM based microcontroller with USB support to perform flash operations. In addition, the microcontroller code can be extended with application code to extend the functionality of the system.

The second thing to take note of is that FPGAs generally have a number of different voltage rails that must be supplied. The FPGA core generally runs at a fairly low voltage - here 1.2V. However, FPGAs have separate supply pins for the IO voltage level (VCCIO), which can be set differently per IO bank if you need to interface with multiple devices at different voltage levels. In our case, we’re going to use 3.3V as our IO voltage, since it is fairly widely compatible and is also the voltage for our microcontroller.

Third, one of the biggest advantages of routing FPGA designs is the lack of constraints around pinout - since any IO pin can be any function, we are fairly free to shuffle our connections between FPGA and PMOD to achieve more direct routing. However, there are some small exceptions - certain pins on the device will be listed as being directly connected to ‘global buffers’, or other special purpose routing constructs that provide high signal fanout / low timing skew within the device itself. If you have any signals coming into the FPGA that are likely to be used in many places (e.g. a clock), it is generally advantageous to connect those signals to a hardware pin that has a dedicated connection to one of the high-fanout internal buffers.

Finished design. 2 Layers.

Building Your Owm

If you want to build your own copy of this breakout board, you will first need to acquire some blank PCBs. If you have done this before, you can clone the repo containing all of the design files (gerbers are in hardware/gerber/) and ship the Gerbers to your preferred PCB manufacturer. Alternately, or for convenience, you can order this design directly from PCBWay, a budget PCB house with respectable quality and turnaround time.

While you wait for your boards, you may also need to order components. The KiCad design files are fully specified with the manufacturer and DigiKey part numbers for all components involved. Also included in the hardware/ folder is a DigiKey BOM (digikey_bom.csv), which can be uploaded to the DigiKey BOM Tool and used to immediately add the necessary components to an order. Note that the BOM item quantity is for precisely one unit - I would recommend going through and adding whatever fudge factor you feel is reasonable to cheaper passives and other parts you want some margin for error with.

At time of writing, the total cost of the BOM for no more than one board is $37.60. If you intend to make more than one, the unit price will go somewhat as you hit the price breaks.

Overview of BOM in KiCad Table Editor

Once you have the parts and boards, there are a couple other prerequisites for assembly -

One of the design goals for this board was for it to be feasible to assemble by hand, using only a soldering iron. To that end, all of the critical components are leaded packages, and there are no passives smaller than 0805. For some components (USB type-C connector J4, 12MHz oscillator X1 and RGB LEDs D1 and U4), a hot air gun may be advantageous, but it is possible to get away without one.

Component placement for assembly

If like me you use a fair amount of flux during assembly, I would recommend washing the PCB afterwards using first some flux cleaner and an ESD toothbrush, followed by a rinse in some 99% isopropyl alcohol. You can then leave the board to air dry, or apply (gentle!) heat with the hot air gun to evaporate any remaining isopropyl.

Finished Assembly (Front)

Finished Assembly (Back)

Interfacing to the Computer

Now that we have our board assembled, we need to be able to write bitstreams to it, otherwise it isn’t very useful. In order to do this, we’re going to need to write some firmware for the onboard microcontroller that will allow us to perform a programming flow along the lines of

For that we are going to use a STM32F0 series part and the open source libopencm3 HAL. For more information on getting started with that programming environment, I would recommend reading through this series of posts.

The core of our firmware is going to be listening for incoming packets over one USB endpoint, figuring out what operation is intended, and then potentially responding back to the computer. The full implementation can be seen in usb_protocol.cpp, but that basic idea is that we use the first byte of any incoming USB packet to determine the command. Since we have a bounded number of commands, we have a lookup table that maps the command code to an appropriate handler function. For example, here is the handler for the UsbProto::Opcode::FPGA_RESET_ASSERT command, which puts the FPGA into reset in anticipation of programming:

void UsbProto::handle_fpga_reset_assert(const uint8_t *buf, int len) {
  // In order to program the flash, we want to make sure that we aren't in
  // contention with the FPGA for control over the SPI lines. The first thing
  // to do here is therefore to pull the NRESET line of the FPGA low
  // Once the FPGA is no longer driving any signals, we can enable the hardware
  // SPI on our microcontroller, taking control of the bus.
  // Once the spi is ready, we send a few initialization commands to the flash
  // chip to ensure it is ready to perform reads / writes.

The vast majority of this code should apply to other boards designed around a microcontroller programming interface in this way, so in an attempt to make this firmware easy to port to your own projects, all pin configuration data is contained in a single header file, include/config.hpp, so that it can be adapted for slightly different implementations.

To get the firmware uploaded, you will need some sort of SWD adapter (my go-to being the BlackMagic Debug Probe), and a TagConnect 2050 cable in order to connect to the programming pads. To build:

# Install ARM toolchain
sudo apt install gcc-arm-none-eabi gdb-arm-none-eabi binutils-arm-none-eabi
# Clone & build the firmware repo
git clone git@github.com:rschlaikjer/hx4k-pmod.git
cd hx4k-pmod
git submodule init
git submodule update
# Build
mkdir programmer_firmware/build
cd programmer_firmware/build
cmake ../
make -j$(nproc)

Once the binary is built, if you are using the black magic probe (and it is connected as /dev/ttyACM0), you can flash the device by first connecting it to power through either the USB header or auxiliary power jack, and then building using the CMake target fpga_programmer_flash. You should see output similar to the following:

ross@mjolnir:/h/r/P/G/f/p/build$ make fpga_programmer_flash
[ 44%] Built target libopencm3
[100%] Built target fpga_programmer_elf
Black Magic Probe (Firmware v1.6.1-311-gfbf1963) (Hardware Version 3)

Available Targets:
No. Att Driver
 1      STM32F07 M0
event_loop () at /home/ross/Programming/Github/fpga-swe-1/programmer_firmware/src/main.cpp:87
Loading section .text, size 0x29e4 lma 0x8000000
Loading section .init_array, size 0x8 lma 0x80029e4
Loading section .data, size 0x14 lma 0x80029ec
Start address 0x8001e84, load size 10752
Transfer rate: 18 KB/sec, 768 bytes/write.
Section .text, range 0x8000000 -- 0x80029e4: matched.
Section .init_array, range 0x80029e4 -- 0x80029ec: matched.
Section .data, range 0x80029ec -- 0x8002a00: matched.
Kill the program being debugged? (y or n) [answered Y; input not from terminal]
[100%] Built target fpga_programmer_flash

You should then see the board show up as a USB device, with the default VID:PID set to a generic test PID:

ross@mjolnir:/h/r/P/G/f/p/build$ lsusb -d 1209:0001
Bus 001 Device 106: ID 1209:0001 Generic pid.codes Test PID

However, this firmware is only one half of the programming flow - we also need a program to run on the host computer and send the necessary control/data packets to the device. To that end, I have created a tool to Find and Flash FPGAs called faff, which speaks the simple USB protocol we’ve defined in the firmware. The implementation of the host side of the USB commands can be seen in src/usb_protocol.cpp, and follows the same naming as used in the firmware, Requests are by and large fairly straightforward - for example, here is the implementation for writing data to flash:

void Session::cmd_flash_write(uint32_t addr, const uint8_t *data,
                              uint8_t size) {
  // Our packet structure must contain
  // - One byte operation (FLASH_WRITE)
  // - Four byte write offset (MSB first)
  // - One byte indicating how many data bytes follow
  // - N data bytes to be written to flash
  uint8_t cmd_out[size + 6] = {
      ((uint8_t)(addr >> 24)),
      ((uint8_t)(addr >> 16)),
      ((uint8_t)(addr >> 8)),
      ((uint8_t)(addr >> 0)),
  // Copy the data to be sent into our output packet buffer
  memcpy(&cmd_out[6], data, size);

  // Ask libusb to send this packet to the device.
  int transferred = 0;
  int ret =
      libusb_bulk_transfer(_usb_handle, _args._usb_endpoint_tx, cmd_out,
                           sizeof(cmd_out), &transferred, libusb_timeout_ms);

  // If libusb failed to write out this packet, we will abort here.
  assert_libusb_ok(ret, "Failed to initiate flash write");

To build this tool, make sure you have libusb’s development headers installed, and then build it using CMake:

# Install dependencies
sudo apt install -y build-essential cmake libusb-dev
# Clone the repo
git clone git@github.com:rschlaikjer/faff.git
# Build as a normal CMake project
mkdir faff/build
cd faff/build
cmake ../
make -j$(nproc)
# Optionally
sudo cp faff /usr/local/bin

Once faff is installed and in your $PATH, we can use it to list connected targets using the -e|--enumerate option, and then load a file using -f:

ross@mjolnir:/h/r/P/G/f/build$ ./faff -e
Searching for devices with VID:PID 1209:0001
[0] Serial: 004700254753511120303234
Found 1 devices

ross@mjolnir:/h/r/P/G/f/build$ ./faff --usb-serial 004700254753511120303234 -f top.bin
Claimed device 1209:0001 with serial 004700254753511120303234
Flash chip mfgr: 0xef, Device ID: 0x17 Unique ID: 0xe4682c404b163333
Programming block 0x00020fa0 / 0x00020fbc
Reading block 0x00020fa0 / 0x00020fbc

faff has a couple more options, which can be found by running it with the --help flag. But for now, being able to detect and write files to the flash is all we need.

Synthesizing Gateware

Finally, we have gotten to the point where we can actually start on the FPGA part of this FPGA breakout board. If you haven’t already, download and install the open source FPGA toolchain consisting of yosys, icestorm (since this is an ice40 part) and nextpnr. Instructions for all of them are on their respective homepages.

The first thing we need to do when bringing up an FPGA board is define the pin mapping. For this, we need a ‘Physical Constraint File’ (pcf). This is something you generally only need to do once per hardware design, and will then be applicable for any applications running on that hardware. The syntax is fairly straightforward (full file for this board can be found here), and assigns a name to a given FPGA pin (by number, for QFx packages, or by ball, for BGA packages):

# Input 12MHz oscillator is connected to physical pin number 21
set_io -nowarn CLK_12MHZ 21

# RGB led has three connections, on FPGA pin numbers 83-85
# Also, note that these LEDs are active LOW!
set_io -nowarn LED_R        84
set_io -nowarn LED_G        85
set_io -nowarn LED_B        83

When we then create our top-level Verilog module, we are able to use any signal defined in the PCF file as a module connection, like so:

`default_nettype none
`define CLK_HZ 12_000_000
module top(
        // Input clock
        input wire CLK_12MHZ,

        // RGB LEDs
        output wire LED_R,
        output wire LED_G,
        output wire LED_B

    // To blink the LEDs at a human scale, let's have a counter that
    // rolls over at approximately 4Hz. Note that this is not a proper
    // prescaler - we're just getting a power of two in the right ballpark
    localparam PRESCALER_4HZ = (`CLK_HZ / 4) - 1;
    reg [$clog2(PRESCALER_4HZ):0] counter;

    // And assign the LED outputs to be the higher (slower) bits of
    // that counter
    assign LED_R = counter[$clog2(PRESCALER_4HZ)];
    assign LED_G = counter[$clog2(PRESCALER_4HZ)-1];
    assign LED_B = counter[$clog2(PRESCALER_4HZ)-2];

    // At every rising clock edge, increment that counter by one
    always @(posedge CLK_12MHZ) begin
        counter <= counter + 1;


If you cd to the gateware/ directory in the repo that accompanies this project and run make, you should see it synthesize, place and pack a demo bitstream that will blink some LEDs as a sign of life. If all goes well, you should see this output:

icepack top.asc top.bin
// Reading input .asc file..
// Reading 8k chipdb file..
// Creating timing netlist..
// Timing estimate: 5.01 ns (199.63 MHz)
// Checking 20.83 ns (48.00 MHz) clock constraint: PASSED.

Now that you have a top.bin file, you can use faff to load it onto the board using the make prog target (assuming faff is in your $PATH):

ross@mjolnir:/h/r/P/G/f/gateware$ make prog
faff -f top.bin
Claimed device 1209:0001 with serial 004700254753511120303234
Flash chip mfgr: 0xef, Device ID: 0x17 Unique ID: 0xe4682c404b163333
Programming block 0x00020fa0 / 0x00020fbc
Reading block 0x00020fa0 / 0x00020fbc

If all goes well, you should be greeted by a green light on the MCU status LED, and a multicoloured blink pattern from the FPGA:

Demo gateware blinking

The demo verilog provided also instructs the FPGA to echo back any UART signals it receives. To verify that this is working properly, connect to the serial port of the board (here ttyACM2) and try typing some characters. You should see them show up in the serial console as they are received, echoed by the FPGA and sent back over USB. You can connect using screen as a serial monitor with the command screen /dev/ttyACM2 2000000, where 2000000 is the baudrate.

Hopefully some of the information or tooling here is useful for those looking to get their feet wet programming or designing FPGA based boards. If you end up using faff or the associated firmware and there is a feature you think would be generally useful, feel free to mention it on either the faff repo or the hardware/firmware repo issues tab.

For a larger project based on this breakout board, take a look at either this post that works through creating an animated GIF display panel using an RGB LED matrix, or this post for creating your own SoC with a RISC-V CPU.

FPGAs for Software Engineers 0: The Basics

Companion code for this post available on Github

In this (increasingly lengthy) post, I will attempt to cover enough of a grounding in what an FPGA is and how it is configured that others out there can start to play around with writing HDL of their own. Be warned that I am by no means an expert on FPGAs, and so information here is only correct to the best of my own knowledge! There is no required hardware to follow along with this post. All tools are open source. I will assume some familiarity with C build tools.

What is an FPGA?

An FPGA (Field Programmable Gate Array) is, first and foremost, NOT A CPU (out of the box, at least). Rather than being a fixed silicon implementation that can only execute what it was made for (like a 74 series logic chip, or an Intel CPU, or an ARM based microcontroller), an FPGA is a large array of basic building blocks, each of which generally implements a very simple truth table. However, on their own these individual blocks would be fairly useless. In order to combine them to form more complex logic, FPGAs also contain a dense set of programmable connections between these logic elements, generally referred to as the routing fabric. If you envision each logic element as a single chip, the routing fabric is a set of programmable wires between all of them. FPGAs can have between hundreds and hundreds of thousands of these individual logic elements, and it is through the combination of many of these blocks that complex logic like conventional CPUs and dedicated special-purpose computation engines can be constructed.

Let’s take a look at a current FPGA, the Lattice ECP5. The figure below is taken from the Lattice ECP5 family datasheet, and shows the Programmable Functional Unit, one of the lower level blocks that make up the FPGA fabric.

ECP5 Programmable Functional Unit (PFU) block

We can see here the two main building blocks within an FPGA - the LUT (Look Up Table) and the Flip Flop (FF). In this case, we have a LUT4 - a lookup table with 4 inputs. With one LUT4, you have 2⁴ = 16 possible states that you can encode. This means that you can trivially perform calculations like AND, OR, XOR, NOT (with up to 4 1-bit inputs) with a single LUT by loading in the appropriate truth table. You are of course not limited to those common boolean operations - a LUT can implement any truth table you may want to create. If it helps to think about FPGAs as compared to discrete logic ICs, an FPGA is essentially hundreds of thousands of individual 74xx logic chips, waiting to be connected together. Thus, anything you could create by hand with discrete logic can also be represented inside the FPGA, without having to breadboard and jumper wire each individual logic element by hand. The other core primitive here is the Flip-Flop, usually of the D variety. These are necessary for buffering outputs of combinatorial logic (LUTs), or as storage for small data (e.g. registers) that you may use in your design.

In addition to the LUT and FF basic elements, many FPGAs also contain some amount of random access memory (RAM), usually of 2-3 distinct classes:

Finally, FPGAs may implement some other special purpose logic elements that can be tied into your design. These include Multiply/Accumulate (MAC) elements, Digital Signal Processing (DSP) elements or full peripherals such as SPIs. These blocks can vary significantly between different FPGA families, so refer to the relevant datasheet for more details on these.

Why would you use an FPGA?

Given the wide and varied offerings of increasingly cheap microcontrollers on the market, you may wonder why you might want to use an FPGA in a project. There are after all negatives to FPGAs - they come with next to nothing built-in. If you get a new microcontroller, you already have a CPU core, peripherals (GPIO, I2C, SPI, etc.) and an interconnect bus tying it all together. With a few bytes of assembly instructions you can write some peripheral registers, and you’re good to go. With an FPGA, at first power on you have absolutely nothing. If you want a CPU, you have to buy, download or build the HDL for one from scratch. You need to implement any connections between your CPU and the peripherals (which, again, YOU have to build into the design).

However, the power of the FPGA is in the fact that you can do these things. If you have a microcontroller, and realize you need another I2C master, you need to go out and buy a different chip. If you discover that your SPI communication is experiencing clock skew, you have to feed the reflected clock back into another SPI peripheral to account for it, which might not be possible or might again require changing out your hardware. But with an FPGA, if you need a new I2C master, you just generate one. If you need to tweak the details of how your SPI master works, you just change the HDL that describes it. Add more FIFO space, clock the MISO line from a reflected CLK signal, whatever you want to change is within your power to change, provided you have the LUTs to do so. This makes FPGAs very powerful for applications where you need to do something specific and do it fast, such as

Hardware Description Languages

What they are not

Now that we have somewhat of a handle on what an FPGA is and when we might want to use one, let’s take a look at how we actually “program” an FPGA. The first thing that I would say to anyone coming to FPGAs from a software background is that:

FPGA Hardware Description Languages are NOT programming languages

This is a very important thing to un-learn, otherwise you may struggle to intuitively understand what is happening in HDLs such as Verilog or VHDL. A programming language like C essentially defines a set of steps that are serial by default. For example, in this code:

int a = 3;
int b = a + 2;
return a + b;

As a programmer, you understand that we first assign 3 to a. Only after we have finished assigning 3 to a do we calculate b, and only after that do we add it back to a and return. This makes sense, since the interface CPUs present to the end user is that they run one instruction, then the next, and so on (even if under the hood, they are speculatively executing and reordering your instructions).

However, on an FPGA, all operations can be thought of as parallel by default. Let’s take a look at the following Verilog:

module foo(
    input wire [7:0] in_a,
    input wire [7:0] in_b,
    output wire [7:0] o_ret1,
    output wire [7:0] o_ret2);

    assign o_ret1 = in_a + in_b;
    assign o_ret2 = (in_a * 2) + in_b;


In this module, the calculations for o_ret1 and o_ret2 happen continuously. This is to say that whenever the input values change, the output also updates, as fast as it is physically possible for the FPGA to do so. Since each statement will instantiate it’s own hardware, these two calculations will also run parallel to each other, such that o_ret1 and o_ret2 will both update at the same time (plus or minus variance in the propagation delay of their logic paths), rather than one after the other. In general, any statement in verilog that stands on its own will be encoded as it’s own logic, and so happen in parallel with other statements. Since what’s happening is not that your code is taking turns to execute on a shared hardware resource (CPU), your code is defining the hardware resources that calculate the desired results. And this points at the crux of the difference:

Programming Languages define a series of steps.

Hardware Description Languages define how to build a circuit.

Not internalizing this difference early on led me personally to spend a lot of time struggling to understand HDLs. Don’t make my mistake!

What they are

As stated before, Hardware Description Languages (HDLs) define circuits. In Verilog, the two basic elements are reg and wire:

Groups of registers and wires can be encapsulated into Modules, which can then be used as their own primitives. This allows us to quickly build up complex behaviours from progressively smaller primitives. A single clocked flip-flop can be repeated to create a FIFO, which can be included as part of a peripheral, several of which can be connected to the same bus for a CPU to access.

What assignment means

In Verilog, there exist three assignment operators:

Blocking assignment is primarily useful for combinatorial / unclocked logic. It allows you to write logic that looks more like a classical program - for example, the following defines a simple combinatorial logic setup:

module blocking(input wire a, input wire b, input wire c, output wire d);
    wire a_xor_b = a ^ b;
    assign d = c & a_xor_b;

Non-blocking assignments are what make sequential logic possible, because of the two-step evaluation. The two time quanta represent the time just before and the time just after the trigger event for the block containing the logic (a positive clock edge, for example). If we consider the following simple divide by 2 oscillator design we can see why this is important:

module osc(input wire i_clk, output reg o_clk);
    always @(posedge i_clk)
        o_clk <= ~o_clk;

If we were to use a blocking assignment here, assigning o_clk to itself would be unresolvable! However, with the two time-steps of the non-blocking assignment this allows triggered logic to function properly - on the first time quanta the new value (the inversion of o_clk) is calculated, and then only on the second is it updated. It is then stable until the next update event.

Combinatorial Logic

Using just our knowledge of wires and assignments, let’s see if we can create a module that will add two (one-bit) numbers together, in addition to a carry-in and carry-out bit. This is known as a Full Adder.

//// A full-adder circuit with combinatorial logic
// We start by defining our new module, with the name 'full_adder'
module full_adder(
    // We have three inputs to our module, the two bits to be added
    // as well as the carry-in value.
    // By default, a wire is one bit wide.
    input wire in_a,
    input wire in_b,
    input wire in_carry,
    // We will output two things: the sum result, as well as whether
    // the sum overflowed (the carry-out).
    output wire out_sum,
    output wire out_carry

// The truth table for a fill adder is like so:
// A B Cin | Sum Cout
// --------|--------
// 0 0  0  |  0   0
// 1 0  0  |  1   0
// 0 1  0  |  1   0
// 1 1  0  |  0   1
// 0 0  1  |  1   0
// 1 0  1  |  0   1
// 0 1  1  |  0   1
// 1 1  1  |  1   1

// Based on the truth table, we can represent the sum as the XOR
// of all three input signals.
assign out_sum = in_a ^ in_b ^ in_carry;

// The carry bit is set if either
// - Both A and B are set
// - The Carry in is set and either A or B is set
assign out_carry = (in_a & in_b) | (in_carry & (in_a ^ in_b));


To see what exactly this verilog will synthesize to, we can use Yosys to read the verilog and generate a flowchart showing each generated primitive and the connections between them. The yosys invocation I use for this is like so:

yosys -q << EOF
read_verilog full_adder.v; // Read in our verilog file
hierarchy -check; // Check and expand the design hierarchy
proc; opt; // Convert all processes in the design with logic, then optimize
fsm; opt;  // Extract and optimize the finite state machines in the design
show full_adder; // Generate and display a graphviz output of the full_adder module

This command on the module above should result in an output that looks a bit like this:

Synthesized Full Adder

The inputs/outputs of our module can be seen in the octagonal labels. From there we can follow the connections to the lower-level blocks that yosys has used to implement our design. If we follow the out_sum label backwards, we can see that the in_carry value is XORed with the output of the XOR of the in_a and in_b signals, which satisfies the wire assignment we made above in the module definition. Similarly, we can see the logic blocks that make up the calculation of out_carry. Notably, we have no buffers in this design - all of these wire assignments are continuous assignments. out_carry and out_sum will update as fast as possible when any of in_a, in_b or in_carry change, bounded only by the propagation delay of the FPGA you might run this on.

Sequential Logic

As with discrete electronics, one doesn’t always want a purely asynchronous setup. Adding a clock and buffers to a design allow for real hold times for data buses, and avoids potential instabilities during long calculation chains. The same is true on FPGAs - while combinatorial logic is good for small chunks of logic, it is generally the case that you will want to tie it into a sequential design for usage. As an example, let’s take the full adder we wrote above and make a 4-bit synchronous adder with it.

// Clocked adder of 4-bit numbers, with overflow signal
module adder_4b(
    // Input clock. At the positive edge of this clock, we will update the
    // value of our output data register.
    input wire i_clk,
    // Input 4-bit words to add. Syntax for multi-bit signals is
    // [num_bits-1:0] (so a 4 bit signal is [3:0]). Individual bits or
    // groups of bits an be accessed using similar notation - signal[1] is
    // the 1st it of that signal. signal[2:1] would be a two-bit slice
    // of signal.
    input wire [3:0] in_a,
    input wire [3:0] in_b,
    // Output data buffer containing the result of the addition.
    // Note that this is a reg, not a wire.
    output reg [3:0] out_sum,
    // Single output bit indicating if the addition overflowed
    output reg out_overflow

// Define two internal wires. We need these to connect the full adder
// elements together as a bus that we can then load into our output buffer.
wire [3:0] full_adder_sum;
wire [3:0] full_adder_carry_outs;

// We now need 4 full adders, with the carry out of each one fed into
// the carry in of the next.
// We could write out all of these by hand, but there exists a verilog
// keyword 'generate', which makes repeated declaration of elements like
// this a lot simpler.

// Because our first adder is special, we need to keep it outside the loop
full_adder adder_0(
    // Take the 0'th bit of the input words as operands
    in_a[0], in_b[0],
    // Since this is the first bit, the carry-in can be hard-coded as 0.
    // The output signals can be connected into the bus we made above so
    // that they can be used by the next adder

// Now we can generate the rest of the adders.
// First, we need to define a loop variable for the generate, like
// we would in any programming language
genvar i;
// Counting from 1 to the total number of bits in the operands, generate
// a new full adder that operates on that bit of the input words, plus
// the carry of the prevoius adder.
// We use the function $bits to get the width of our in_a parameter, so
// that if it changes we don't need to modify this code.
for (i = 1; i < $bits(in_a); i++)
    full_adder adder(
        // Add the i'th bit of both operands
        // Use the carry-in from the previous adder
        // Connect our sum output to the full adder bus
        // Likewise connect our carry out so that it can be used by
        // the next full adder

// Now we can set up the sequential logic part of this adder.
// For clock or other event-driven signals, we need a sensitivity
// selector. Whenever the conditions in parenthesis are met, the
// body of the block will execute.
always @(posedge i_clk) begin
    // At the positive clock edge, we want to take the current value
    // that is being calculated using combinatorial logic by the full
    // adder, and copy that data into our output data buffer.
    // For this, we use the non-blocking assignment operator, <=
    out_sum <= full_adder_sum;
    // Similarly, we take the final carry bit of the adder and use
    // that as our overflow indicator.
    out_overflow <= full_adder_carry_outs[$bits(in_a)-1];


Now that we have defined our 4 bit adder, let’s open it in a flowchart the same was as we did for our single bit full adder. Note that you may need to include the full_adder module definition (or a dummy implementation) in the same file as your 4-bit adder for yosys to synthesize.

Clocked 4-Bit Adder

There is a lot more going on here! On the far left we can see our initial full adder, with the zero’th bits of the input operands and a literal 0 bit as inputs. The outputs then go off to the wire buses we defined, and we can see the block of generated adder modules in the center. At the end, the final carry bit goes to a D flip flop, which is clocked by our i_clk signal. Similarly, the output sum data is fed to a 4-bit DFF that is also clocked.

Simulating with Verilator

In order to test and debug our HDL code, we could deploy it directly to a physical FPGA and hope for the best, but it can be much easier to first simulate parts or all of the design on the computer, allowing us to control and observe signals at a very low level, and to trace execution of the HDL over an arbitrary number of clock cycles. For Verilog, there exists an open source tool called Verilator that can take your Verilog source code and convert it to ordinary C code, which can then be linked against test benches of your own design. Let’s verify that our 4-bit adder works properly by creating a test bench that tests every possible combination of 4-bit values, and asserts that the result is correct.

First, we’re going to create a CMake based verilator setup (supported only on more recent versions of verilator, I recommend at least 4.023). In our CMakeLists.txt, let’s add the following:

cmake_minimum_required(VERSION 3.8)

# We need to locate the verilator install path.
# If this doesn't work off the bat on your system, try setting
# the VERILATOR_ROOT environment variable to your install location.
# On a default install, this will be /usr/local/share/verilator
find_package(verilator HINTS $ENV{VERILATOR_ROOT} ${VERILATOR_ROOT})
if (NOT verilator_FOUND)
    message(FATAL_ERROR "Could not find Verilator. Install or set $VERILATOR_ROOT")

# Create our testbench executable.
# We will have just the one testbench source file.
add_executable(adder_testbench adder_testbench.cpp)

# Based on how I understand the verilator plugin to work, if you specify several
# verilog sources in the same target it will pick a top level module and
# only export the interface for that module. If you want to be able to poke
# around in lower-level modules, it seems that you need to specify them
# individually. To make this easier, I put all the sources in a list here
# so that I can add targets for them iteratively.

# For each of the verilog sources defined above, create a verilate target
# We need to speicify an existing executable target to add the verilog sources
# and includes to. We also speicfy here that we want to build support
# for signal tracing
verilate(adder_testbench COVERAGE TRACE
    VERILATOR_ARGS -O2 -x-assign 0

Now that we have a framework to build our testbench, let’s write something fairly straightforward in C. We will manually toggle the clock line while writing all our test inputs to the adder. We will then take the output and assert it matches what we calculate with C.

#include <stdlib.h>
#include <verilated.h>
#include <verilated_vcd_c.h>

#include <Vadder_4b.h>
#include <Vfull_adder.h>

int main(int argc, char **argv) {
  // Initialize Verilator
  Verilated::commandArgs(argc, argv);

  // Enable trace

  // Create our trace output
  VerilatedVcdC *vcd_trace = new VerilatedVcdC();

  // Create an instance of our module under test, in this case the 4-bit adder
  Vadder_4b *adder = new Vadder_4b();

  // Trace all of the adder signals for the duration of the run
  // 99 here is the maximum trace depth
  adder->trace(vcd_trace, 99);

  // Output the trace file

  // We need to keep track of what the time is for the trace file.
  // We will increment this every time we toggle the clock
  uint64_t trace_tick = 0;

  // For each possible 4-bit input a
  for (unsigned in_a = 0; in_a < (1 << 4); in_a++) {
    // For each possible 4-bit input b
    for (unsigned in_b = 0; in_b < (1 << 4); in_b++) {
      // Negative edge of the clock
      adder->i_clk = 0;
      // During the low clock period, set the input data for the adder
      adder->in_a = in_a;
      adder->in_b = in_b;
      // Evaluate any changes triggered by the falling edge
      // This includes the combinatorial logic in our design
      // Dump the current state of all signals to the trace file

      // Positive edge of the clock
      adder->i_clk = 1;
      // Evaluate any changes triggered by the falling edge
      // Dump the current state of all signals to the trace file

      // The adder should now have updated to show the new data on the output
      // buffer. Assert that the value and the overflow flag are correct.
      const unsigned expected = in_a + in_b;
      const unsigned expected_4bit = expected & 0b1111;
      const bool expected_carry = expected > 0b1111;
      // Check the sum is correct
      if (adder->out_sum != expected_4bit) {
        fprintf(stderr, "Bad result: %u + %u should be %u, got %u\n", in_a,
                in_b, expected, adder->out_sum);
      // Check the carry is correct
      if (expected_carry != adder->out_overflow) {
                "Bad result: %u + %u should set carry flag to %u, got %u\n",
                in_a, in_b, expected_carry, adder->out_overflow);

  // Flush the trace data
  // Close the trace
  // Testbench complete

Now, if we build and run our testbench like so:

mkdir build
cd build
cmake ../

We should see… nothing! Our asserts should pass, and the testbench won’t error out. We should also see that in the same directory from which we ran the testbench, an adder_testbench.vcd file will have been created. This contains the data of all the signals during our test case. We can open this file in GTKWave or another wave viewer. If we click on the TOP block in the SST tree, we should see the signals that are the inputs and outputs to our four bit adder appear in the list view below. Select all of them and hit ‘Append’ to place these signals on the scope view to the right. Now you can scroll through time and observe that at the negative edge of each clock, we set the new input data for the adder (this could also be done at the posedge. At the following positive edge, we can see that the output of the adder module correctly updates with the sum and overflow value.

Trace output for our 4-bit adder testbench

Hopefully with this foundation, you feel comfortable writing, viewing the generated logic for and testing your own simple logic blocks. The full source code for the snippets listed in this post can be found on Github. If you have any corrections or improvements to suggest, feel free to do so.

In the next post, we create and deploy to an FPGA breakout board based on the Lattice Ice40 HX4k. After that, we use it to create an Animated GIF display on an LED panel and a RISC-V based SoC.

STM32 & OpenCM3 5: Debugging & Fault Handlers

Companion code for this post available on Github

This is the sixth post in a series on the STM32 series of MCUs. The previous post, on memory sections and ITCM, can be found here.

When building embedded software, it can sometimes be challenging to determine the root cause of a failure. You have no operating system to fall back on, and if you are coming from a non-embedded background may not know about some of the low-level behaviours that can be leveraged to make finding certain bugs a little easier. In this post, I’ll go over in detail the steps to connect an interactive debugger to your embedded system, as well as how you can make interrupt problems easier to diagnose.

Connecting using GDB

If you have worked on applications software in either C or C++ for some time, here’s a good chance you will at least have heard of GDB. As one of the more venerable debuggers, it supports an incredible number of targets, including (to my knowledge) the entire ARM Cortex ecosystem.

There are a plethora of ARM programming and debug adapters on the market, such as:

For the purposes of this post, we will be using the cheap and powerful Black Magic Probe (BMP). This probe comes with a standard 10-pin SWD header, and a 4-pin UART connector. Both of these have level-shifted output between 1.7 and 5V, and the probe can also supply 3.3V power to targets if necessary. Even better, the probe runs an implementation of the GDB server protocol, meaning that it can be connected to GDB with no additional software dependencies, unlike some other tools.

To get started with the black magic probe, you’ll first need to connect it to your device under test. For this you can use the included 0.127mm pitch SWD connector, or other standard cable such as the popular TagConnect TC-2050 cable, which eliminates the need for costly headers.

Two example connections of the probe, one with the included SWD cable, and one with a TagConnect cable, are shown here:

Example connections of Black Magic Probe

Once your cable is connected, and the device is powered up, it’s time to fire up GDB. If you don’t already have it, you can install it and the rest of the GNU ARM toolchain on Debian based systems with the following command:

sudo apt install binutils-arm-none-eabi gcc-arm-none-eabi gdb-arm-none-eabi

For other systems, you can install GDB by following the instructions here.

Now that we have GDB ready, we need to load up the program we’re debugging and connect to our probe. To do that, we first start GDB with the elf as our first argument:

$ arm-none-eabi-gdb my_program.elf
GNU gdb (7.12-6+9+b2)
Copyright (C) 2016 Free Software Foundation, Inc.
# Connect to the black magic probe as a target. There are two serial ports
# exposed by the device - the first is the GDB server, the second is a
# passthrough to the UART pins on the device. The exact path to the serial
# devices will vary by OS.
(gdb) target extended-remote /dev/ttyACM0
# Now that we have told GDB to use the BMP as a target, we can invoke
# some extra commands. The first will be to scan for serial wire debug targets,
# which should return the MCU that we wish to connect to.
(gdb) monitor swdp scan
# Now that we've scanned and identified the device we want to connect to, we
# can start debugging it by attaching. In this case we want the first (and only)
# target. Invoking the attach command will automatically halt the device and
# show the current stack frame.
(gdb) attach 1

If you find yourself typing in these commands a lot, I’d recommend putting them all in a file called .gdbinit in your working directory. This will cause GDB to automatically run them on startup, saving you some time if you are repeatedly closing and reopening GDB to connect to the same system. Note that you may need to set the configuration value

set auto-load safe-path /

in the .gdbinit in your home directory to allow loading of arbitrary .gdbinit scripts. To be more secure, use a more specific path than /, otherwise malicious source trees could cause you to run arbitrary GDB commands.

Now that we are connected to the device, we can interact with it almost as though it were a program runnig on our local machine, but with a few extra commands. I’ll list a couple things I use frequently here, but do consult the GDB manual both for more details about these commands and for other useful invocations. All commands shown below can be shortened to the character in [].

# Print an expression. Expression can be a variable, function, macro, etc.
> [p]rint myvar / MY_DEBUG_MACRO / *((uint32_t*)0xDEADBEEF)
# Exampine one or more memory addresses.
> e[x]amine my_array
> x/10ub 0xDEADBEEF # Print 10 unsigned bytes starting at 0xDEADBEEF
> x/s 0xDEADBEEF # Print the C-string starting at 0xDEADBEEF
# Breakpoints
> [b]reakpoint main # Create breakpoint on entry of method main()
> b my_code.cpp:123 # Create breakpoint at line 123 of my_code.cpp
> [d]elete 2 # Delete breakpoint number 2
# Memory watchpoints
> watch my_var # Trigger breakpoint if my_var changes
# Trigger breakpoint if the integer value located at 0xDEADBEEF changes
> watch watch *(int *) 0xDEADBEEF
# Control flow
> [n]ext # Advance execution to next line of code
> [ni] (next instruction) # Advance execution to next assembly instruction
> [c]ontinue # Run until breakpoint or user interrupt (ctrl-c)
> [r]un # Start program over from beginning
# Information
> info locals # Print all local variables and their values
> info registers # Print the contents of all CPU registers
> info breakpoints # Print all the currently active breakpoints
# Flashing
> file my_program.elf # Select the active binary
> load # Flash the microcontroller with the active file
> compare-sections # Verify the microcontroller code matches the active file
# Dump memory range 0x0 to 0xFFFF to file out.bin as raw binary data
> dump binary memory out.bin 0x0 0xFFFF

As an added tip, if your program is compiled using a makefile any make commands entered inside of GDB will be run normally, allowing you to rebuild and re-flash your microcontroller without ever leaving your GDB session.

Fault Handlers

While the GDB commands above should handle a lot of debugging needs, there will still be some cases (generally due to interrupts) where the control flow of the CPU is hard to follow. In cases like these, implementation of the ARM hard fault interrupt, as well as providing default implementations for all user interrupts, can be very useful.

If your code enables an interrupt but doesn’t implement the associated handler, or causes a processor fault by attempting to execute an invalid instruction or access invalid memory, the ARM core will jump to the Hard Fault Handler, which is an interrupt common to the entire Cortex-M family. Implementing this interrupt handler, and using it to provide error feedback, can save many hours of second guessing your code.

Generally, your HAL library (such as CMSIS, or in this example series, libopencm3) will provide a weakly linked implementation of these system interrupt handlers. In order to provide your own, all you need to do is implement it, which will override the weakly linked one in the HAL.

When you do implement the hardfault handler, the first thing you will want to do is determine which stack pointer was in use when the program crashed (this is mainly relevant for applications making use of an RTOS or other system that takes advantage of multiple hardware stacks). When an exception or interrupt handler is entered, the processor updates the link register with a special value, EXC_RETURN. The full details of this behaviour can be found in the ARMv7-M reference manual section B1.5.8, but for our purposes the salient bit is bit 2, which determines whether the return stack is the main stack, or a process stack. By testing the link register against the pattern (1 << 2), we can determine which stack point was in use when the exception occurred, and pass the appropriate one through to our generalized exception handler.

I’ve seen a couple hardfault handler implementations online that require a CPU supporting conditional execution, which isn’t present on the Cortex-M0 series processors. Here’s a fault handler that should be generic enough to work on all ARMv7-M processors, at the cost of a few extra instructions:

void hard_fault_handler(void) {
    "MRS r0, MSP\n" // Default to the Main Stack Pointer
    "MOV r1, lr\n"  // Load the current link register value
    "MOVS r2, #4\n" // Load constant 4
    "TST r1, r2\n"  // Test whether we are in master or thread mode
    "BEQ base_fault_handler\n" // If in master mode, MSP is correct.
    "MRS r0, PSP\n" // If we weren't in master mode, load PSP instead
    "B base_fault_handler"); // Jump to the fault handler.

With this bridge method in place, we can write the meat of our fault handler code. We take as input the stack pointer address that was determined by our assembly bridge, and gather some pertinent information about the crash into local variables for inspection.

// Core ARM interrupt names. These interrupts are the same across the family.
static const char *system_interrupt_names[16] = {
    "SP_Main",      "Reset",    "NMI",        "Hard Fault",
    "MemManage",    "BusFault", "UsageFault", "Reserved",
    "Reserved",     "Reserved", "Reserved",   "SVCall",
    "DebugMonitor", "Reserved", "PendSV",     "SysTick"};

void base_fault_handler(uint32_t stack[]) {
  // The implementation of these fault handler printf methods will depend on
  // how you have set your microcontroller up for debugging - they can either
  // be semihosting instructions, write data to ITM stimulus ports if you
  // are using a CPU that supports TRACESWO, or maybe write to a dedicated
  // debug UART
  fault_handler_printf("Fault encountered!\n");
  static char buf[64];
  // Get the fault cause. Volatile to prevent compiler elision.
  const volatile uint8_t active_interrupt = arm::scb::ICSR & 0xFF;
  // Interrupt numbers below 16 are core system interrupts, we know their names
  if (active_interrupt < 16) {
    sprintf_(buf, "Cause: %s (%u)\n", system_interrupt_names[active_interrupt],
  } else {
    // External (user) interrupt. Must be looked up in the datasheet specific
    // to this processor / microcontroller.
    sprintf_(buf, "Unimplemented user interrupt %u\n", active_interrupt - 16);

  fault_handler_printf("Saved register state:\n");
  __asm volatile("BKPT #01");
  while (1) {

If you were to have GDB attached to your microcontroller when this handler is hit, you will automatically hit the breakpoint triggered by __asm volatile ("BKPT 01"), and be able to get a summary of what went wrong by asking GDB for info locals, as well as investigate the additional information printed out over our serial console:

GDB info locals command output

In addition to the variables in the above method, we call a dump_registers method to interpret and print the values of the calling stack frame that were saved by the CPU before it jumped to the exception handler. The list of registers, and the order in which they appear, is listed in section B1.5.6 of the ARM reference manual. We can use this info to generate some more debug output, like so:

enum { r0, r1, r2, r3, r12, lr, pc, psr };
void dump_registers(uint32_t stack[]) {
  static char msg[32];
  sprintf_(msg, "r0  = 0x%08x\n", stack[r0]);
  sprintf_(msg, "r1  = 0x%08x\n", stack[r1]);
  sprintf_(msg, "r2  = 0x%08x\n", stack[r2]);
  sprintf_(msg, "r3  = 0x%08x\n", stack[r3]);
  sprintf_(msg, "r12 = 0x%08x\n", stack[r12]);
  sprintf_(msg, "lr  = 0x%08x\n", stack[lr]);
  sprintf_(msg, "pc  = 0x%08x\n", stack[pc]);
  sprintf_(msg, "psr = 0x%08x\n", stack[psr]);

This works well as a generic fault handler, but there are some cases where we may want to also include some additional information. For example, if a memory fault occurs, there are several potential causes that we can flag up, as well as the address at which the fault occurred. So for handling memory faults, we could add a handler function such as the following:

void mem_manage_handler(void) {
  // Pull the MMFSR data out into variables for easy inspection
  // Variables are volatile to prevent compiler elision
  const volatile bool mmfar_valid =
      arm::scb::CFSR & arm::scb::CFSR_MMFSR_MMARVALID;
  const volatile bool fp_lazy_error =
      arm::scb::CFSR & arm::scb::CFSR_MMFSR_MLSPERR;
  const volatile bool exception_entry_error =
      arm::scb::CFSR & arm::scb::CFSR_MMFSR_MSTKERR;
  const volatile bool exception_exit_error =
      arm::scb::CFSR & arm::scb::CFSR_MMFSR_MUNSTKERR;
  const volatile bool data_access_error =
      arm::scb::CFSR & arm::scb::CFSR_MMFSR_DACCVIOL;
  const volatile bool instruction_access_error =
      arm::scb::CFSR & arm::scb::CFSR_MMFSR_IACCVIOL;

  // Pull the MMFAR address
  const volatile uint32_t mmfar_address = arm::scb::MMFAR;

  // Trigger a breakpoint
  __asm volatile("BKPT #01");

If you have an MPU and trigger a write violation, or try and perform operations with an invalid alignment, you will trigger this method and like before can get the problem at a glance with info locals:

Memfault handler locals after null deference

In the output above, we can see that the MemManage Fault Address Register (MMFAR) has been loaded with the address of the error, that the MMFAR address is at 0x0, and that the access type was a data access. In other words, a null pointer dereference!

If you want to have several exception vectors map to the same handler, a useful trick is to alias those methods using an __attribute__ directive like so:

void bus_fault_handler(void) __attribute__((alias("hard_fault_handler")));
void usage_fault_handler(void) __attribute__((alias("hard_fault_handler")));

Keeping things clean

In order to keep my registers organized, I like to nest them in C++ namespaces instead of having a huge list of preprocessor macros (as can be seen in the methods above). In order to make defining registers simple, and still allow the compiler to optimize nicely, I use a template like the one below to generate references to each ARM register:

namespace arm {

// Convenience template for taking an integer register address and converting
// to a reference to that address.
template <typename T> constexpr T &Register(uint32_t addr) {
  return *reinterpret_cast<T *>(addr);

// Typedefs for register references of 32, 16 and 8 bits.
using Reg32 = volatile uint32_t;
using Reg16 = volatile uint16_t;
using Reg8 = volatile uint8_t;

} // namespace arm

Using that template, we can then go through and quickly define each of the registers in the ARM System Control Block (SCB).

namespace arm {
namespace scb {
// Interrupt control and state register (RW)
static Reg32 &ICSR = Register<uint32_t>(0xE000ED04);
// Configurable Fault Status Register
static Reg32 &CFSR = Register<uint32_t>(0xE000ED28);
// MemManage Fault Address Register
static Reg32 &MMFAR = Register<uint32_t>(0xE000ED34);

//// Register subfields
// 1 if MMFAR has valid contents
const uint32_t CFSR_MMFSR_MMARVALID = (1 << 7);
// 1 if fault occurred during FP lazy state preservation
const uint32_t CFSR_MMFSR_MLSPERR = (1 << 5);
// 1 if fault occurred on exception entry
const uint32_t CFSR_MMFSR_MSTKERR = (1 << 4);
// 1 if fauly occurred on exception return
const uint32_t CFSR_MMFSR_MUNSTKERR = (1 << 3);
// 1 if a data access violation occurred
const uint32_t CFSR_MMFSR_DACCVIOL = (1 << 1);
// 1 if an eXecute Never violation has occurred
const uint32_t CFSR_MMFSR_IACCVIOL = (1 << 0);
} // namespace scb
} // namespace arm

With any modern compiler, these constants will be nicely inlined, resulting in zero runtime overhead compared to the old-school #define method. Depending on how you like to write your code, either method will work just as well; this is more a personal preference point than anything else.

Hopefully some of this information comes in useful when debugging your own embedded projects. As ever, a Github repo containg some example code is available here.

DEFCON 27: Badge Writeup

Companion code for this post available on Github

If you just want to cut to the chase and flash your own badge with the Chameleon firmware, grab this build and jump to the “Flashing the badge” section.

This year at DEFCON, we were lucky enough to be provided with another electronic badge, this time courtesy of Joe Grand. The badge is a very sleek design featuring a quartz face, lanyard mounting straps and a Kinetis KL27 series microcontroller (specifically, a KL27P64M48SF2 ). The badge also has an unusual communication mechanism, an NXH2261UK Near-Field Magnetic Induction chipset and antenna.

This year the core badge hardware was the same across badge types, the only differences being a ‘badge type’ byte in the firmware for each badge, and various colours of quartz on the non-human badges.

Human Badge, Front and Back

Some more information on the badge itself, including pictures of all the badge types, can be found on Joe Grand’s Website.

If one wanted to complete the badge without any trickery, they would need to go around the conference interacting with all of the other badge types, including a select few ‘magic’ badges, in order to complete their badge. If we peek the source code of the badge, we can see exactly what’s needed:

// Bit masks for badge quest flags
#define FLAG_0_MASK 0x01 // Any Valid Communication
#define FLAG_1_MASK 0x02 // Talk/Speaker
#define FLAG_2_MASK 0x04 // Village
#define FLAG_3_MASK 0x08 // Contest & Events
#define FLAG_4_MASK 0x10 // Arts & Entertainment
#define FLAG_5_MASK 0x20 // Parties
#define FLAG_6_MASK 0x40 /* Group Chat (all 6 gemstone colors:
                          Human/Contest/Artist/CFP/Uber +
                          Goon + Speaker + Vendor + Press + Village) */

To save yourself some walking and learn a bit more about the badge firmware, read on and we’ll cover two ways to complete the badge the hardware hacking way.


In order to debug or flash your device, you’ll need one of the many ARM programmers available. Joe Grand recommended NXP’s own LPC-Link 2 but you can likely use any debug probe like the Segger J-Link ($$), or the Black Magic Probe (much more affordable), which is what I’ll be using.

You will also need a particular TagConnect cable, the TC-2050-IDC-NL-050-ALL, or some fine gauge wire and a steady hand. If you plan on developing many of your own ARM based designs, I would strongly recommend you pick up the cable. The convenience and cost savings of not having to place .127” pin headers quickly makes up for the price of the cable. You may also want to pick up some cable retaining clips, which make extended debugging require one fewer arms.

If you intend to compile for or flash your badge, you will also need the GCC ARM toolchain, which you can install using your package manager of choice:

sudo apt install binutils-arm-none-eabi gcc-arm-none-eabi gdb-arm-none-eabi

First Approach: No firmware rewriting

Our initial goal was to solve the badge “legit” (for some definition of the word), by not rewriting the firmware in any way. For this method, you will need to populate the 1.8V serial headers on the opposite side to the tag-connect pads. This method will require two badges, and the workflow goes like this:

In order to trick badge A into thinking it’s complete, we first need to figure our what memory location holds the game state variable. I’m not skilled at RE, so instead I’ll cheat a little, and use the linker map, which is available on the DEFCON media server. As a quick recap for those that haven’t seen linker maps before, the map contains the load and virtual address of all variables and functions in the finished binary. It can come in extremely helpful when debugging embedded systems, as we’ll see here. If we search through the map file for the badge_state variable, we can see that it’s located at SRAM address 0x1ffffcdc:

                0x1ffffcdb        0x1 ./source/dc27_badge.o
                0x1ffffcdc        0x1 ./source/dc27_badge.o
                0x1ffffcdd        0x1 ./source/dc27_badge.o

This means that to trick our badge, we just need to overwrite this one memory address. To do that, we’ll connect to the badge over SWD (check the “Flashing the badge” section below for a more thorough explanation of this process) in order to debug it using GDB. Once you’ve attached to the badge, there’s only one necessary command to set the flags:

set {char}0x1ffffcdc = 7

Once you’ve done that, hit c for continue to un-halt the badge CPU. You can now connect the four serial lines of the black magic probe to the UART pinout on the opposite side of the battery. With the badge face down and the SWD connector to the south, the staggered pinout for serial is GND (black), TX (green), RX(purple), VCC (red). Unlike with some of the other programmers, you must connect the power line of the black magic probe in order to power the on-board level shifter. With other programmers, be aware that they may attempt to power the badge themselves, and applying voltages over the expected 1.8V badge voltage may cook your badge.

Now that you have your game state updated, when you connect to the serial console using screen /dev/ttyACM1 115200, you should be greeted with three additional options:

Extended serial commandset on complete badge

The last of these, ‘Update Transmit Packet’, is what we’ll be using to get our B badge to complete. If we take a look at the firmware, we can see how the packets are constructed:

struct packet_of_infamy // data packet for NFMI transfer
  uint32_t uid;   // unique ID
  uint8_t type;   // badge type
  uint8_t magic;  // magic token (1 = enabled)
  uint8_t flags;  // game flags (packed, MSB unused)
  uint8_t unused; // unused

With this info, we can craft our own packets, masquerading as any badge we want. Here’s a complete list of badge types, and the command you need to send in order to become them:

Human   U 772502840001ff00
Goon    U 772502840101ff00
Speaker U 772502840201ff00
Vendor  U 772502840301ff00
Press   U 772502840401ff00
Village U 772502840501ff00
Contest U 772502840601ff00
Artist  U 772502840701ff00
CFP     U 772502840800ff00
Uber    U 772502840901ff00

After manually cycling through these on badge A to unlock badge B, you can then reset badge A and do the same for it, resulting in two ‘legitimately’ unlocked badges. Of course, if that’s not enough for you, we can take it one step further: automating the process by building a Chameleon badge.

Second Approach: Building a Chameleon

After spending several hours manually rotating packets to advance other badges, we decided it was time to automate the process. Since the firmware is freely available on the DEFCON Media Server we can grab it, modify it to our heart’s content and then flash it to our badge. The first stumbling block I hit is that the software is written to rely on NXP’s own libc implementation, Redlib, and downloading the official NXP toolchain on the DEFCON wifi was going to take 8 hours. Instead of that, I rewrote the software slightly to use Newlib, which is packaged along with the GCC ARM toolchain. The full modified firmware building against newlib-nano (and including the chameleon patches) is available in this Github repo.

Once we have the original firmware building, we can go about editing it to broadcast as every other badge type. In order to hook this in, there are two main changes to be made. The first is that we need some way of keeping track of time - the systick implementaion Joe used here only acts as a blocking countdown timer, since he’d just been using it for delays. Since we want to keep an idea of how long we’ve been broadcasting as one badge, we need to add a second counter we can use as monotonic time. This is a quick two line change:

--- a/source/dc27_badge.c
+++ b/source/dc27_badge.c
@@ -423,6 +423,7 @@ static uint32_t pflashSectorSize = 0;

 // Timer
 volatile uint32_t g_systickCounter;
+// Monotonic count-up timer
+volatile uint32_t g_monotonicTime = 0;
 volatile bool g_lptmrFlag;

 // UART2 (to/from host)
@@ -2935,6 +2952,7 @@ void SysTick_Handler(void)
+    g_monotonicTime++;

Now, every time the SysTick interrupt is generated, as well as decrementing the delay timer we will also increment our own monotonic timer. The SysTick timer is configured to interrupt every 1ms, however it is also paused when the badge enters sleep mode, which it does while not actively transmitting / receiving packets. Since our timer will only advance during transmit, we can use a relatively short interval in our code, since it will get stretched out as the chipset sleeps. With the timer in place, the patch for a ‘chameleon’ badge is relatively straightforward, and is added at the top of the while (1) block in main():

// Outside our loop, declare our state variables:
static uint32_t state_change_timer = 0;
// Inside our while(1) loop, handle the chameleon code
if (g_monotonicTime > state_change_timer + 1000) {
  // If it's been 1000 systicks since we last changed our badge state, it's
  // time to update. First, reset our timer to the current monotonicTime.
  state_change_timer = g_monotonicTime;
  // Now, increment our badge type by 1, changing our identity.
  nxhTxPacket.type = nxhTxPacket.type + 1;
  // If we've cycled through all the way to UBER (which is currently read only
  // as Human, and so not particularly useful to broadcast) then go back to the
  // beginning, skipping human and starting at Goon.
  if (nxhTxPacket.type >= UBER) {
    nxhTxPacket.type = GOON;
  // Having changed our packet struct, we now need to load it into the
  // NXH2261 to be broadcast.
  if (KL_UpdatePacket_NXH2261(nxhTxPacket)) {
    // If we fail once, try again. Joe seems to do this elsewhere in the code.
    if (KL_UpdatePacket_NXH2261(nxhTxPacket)) {
      // If we fail a second time, give up and log a message.

There are some other fun things you can do in the firmware, such as enabling a longer version of everyone’s favourite song, or editing your LED pattern, but for the purposes of this post those are left as exercise to the reader. Once you’ve made your mods, from the Firmware/Debug folder you can run make dc27_badge.axf to rebuild the firmware. If all goes well, you should get a nice printout of your memory utilization and a success message:

Memory region         Used Size  Region Size  %age Used
   PROGRAM_FLASH:       58224 B        64 KB     88.84%
            SRAM:        5228 B        16 KB     31.91%
         USB_RAM:          0 GB        512 B      0.00%
Finished building target: dc27_badge.axf

Flashing the badge

Now that we have our updated badge firmware, it’s time to flash. I’ll assume a black magic probe here, for other probes please consult their manuals. The first thing we need to do is update the firmware on our black magic probe. There seems to be a bug in the latest official firmware (at time of writing, 1.6.1) where the KL27x64 series is not recognized properly, and will show up as a generic Cortex-M part. This causes the flashing to fail, since the KL27x64 require a specific flash unlock code before programming. To update your black magic probe, clone the firmware repo from Github build it with make and then perform a DFU update on your probe with the following command:

sudo dfu-util -d 1d50:6018,:6017 -s 0x08002000:leave -D src/blackmagic.bin

Now that the firmware is updated, we can connect to the badge. You will first need to either

Removing the quartz will keep your cable intact, and allow you to clip the cable in place for longer development sessions. The adhesive is strong enough that it can survive being carefully removed and reattached a few times.

Once you have connected the tag-connect cable to the badge one way or the other, it is time to fire up arm-none-eabi-gdb and do our dirty work. Once you have GDB open, we first need to point it to the black magic probe as our remote debugging tool. To do this, we use the command target extended-remote /dev/ttyACM0, where /dev/ttyACM0 should be the first of the two serial endpoints exposed by the black magic probe. Now that we have GDB connected to the probe, we can scan for devices, using monitor swdp scan. This should return a list of two devices for the DEFCON badge: the chipset, and a recovery mode that I have not explored. N.B: If your scan doesn’t return KL27x64 M0+, and instead returns ‘Generic Cortex-M’, close GDB and retry. This seems to be a race condition of some sort. Since we want to debug the main chip, we attach to it using attach 1, which halts the core and prints our current stack frame. A successful attach session should look somewhat like the following:

Successful attach over GDB

Now that we are hooked up and ready to go, flashing the badge is relatively straightforward: we need to select the file to load using file /path/to/dc27_badge.axf, and then to program the badge we just need to run load. If you don’t want to compile your own firmware, you can use this build of a chameleon badge that I’ve created. Once you run load, you should see output like the following:

Firmware flashing over GDB

Once the load is complete, your badge will still be in a halt state. To get it running, either hit c for continue in GDB, or detach the probe and power cycle the badge. You should now have your own chameleon or otherwise custom firmware loaded up!


If this all seemed interesting, don’t be afraid to try it! For some more reading on developing for embedded ARM systems you can check out this tutorial series, and to keep abreast of the progress hacking next year’s badge, join the Hack the Badge slack for discussion.

You can also check out some other writeups of the badge, by some of the great people I met at the HHV this year:

STM32 & OpenCM3 4: Memories and Latency

This is the fifth post in a series on the STM32 series of MCUs. The previous post, on CANbus, can be found here.

As core frequencies increase, the performance penalty of loading instructions and data from slow flash memories increase. For a modern core such as the STM32F750 running at 216MHz, a single flash read can stall the CPU for 8 cycles. Luckily, on these faster cores exist mechanisms to ameliorate or eliminate these stalls. Here I will go over two methods: loading functions into SRAM for zero-wait-state execution, and enabling of the built-in I-cache on certain Cortex-M processors.

Executing from SRAM

Above a certain core frequency, it is no longer possible for the attached flash memory to keep up. This results in the need for ‘wait states’ inside the CPU - in order to progress to the next instruction, the CPU must stall and wait for the fetch from flash to complete. Even worse, the faster your CPU frequency gets, the more pronounced this problem becomes.

A first solution to this, that works on all embedded microcontrollers which allow executing from memory, is to simply copy the code to be executed from flash into SRAM once at startup, and then run it from there afterwards. Since most microcontrollers have single-cycle access latency for SRAM, this can increase execution time significantly. Even on microcontrollers that support some amount of prefetching, copying functions to memory can be useful for code that must execute in deterministic time, or that is frequently jumped to from unpredictable locations (for example, interrupt service routines).

Luckily, there is a way to achieve this with GNU utilities. By default, all code will be placed into the .text section by your compiler. But it doesn’t have to! We can create our own sections, and do as we please with them. So for now, let’s create a section called sram_func and designate a function as being part of this section. First, let’s edit our linker script to tell it where the new section should go, and what it should contain.

/* Presumably, you already have a section like this defining the physical layout
   of your particular microcontroller */
    rom (rx) : ORIGIN = 0x08000000, LENGTH = 64K
    ram (rwx) : ORIGIN = 0x20000000, LENGTH = 320K

/* Other existing directives should likely be kept above this new section,
   unless you already have something fancy going on and know better */

/* Now, we can create a new section definition */
.sram_func : { /* Our new section will be called .sram_func */
    /* This creates a new variable, accessible from our C code, which points
       to the start address of our newly created section, in memory.
       The reason for this will become apparent later. */
    __sram_func_start = .;
    /* We now tell the linker that into the .sram_func section it should place
       all of the section data we will later tag as 'sram_func'. */
    /* Pad the end of our section if necessary to ensure that it is aligned on a
       32-bit word boundary */
    . = ALIGN(4);
    /* We now take the end address of this new section, and make it also
       available to the program. */
    __sram_func_end = .;
} >ram AT>rom /* These two directives control the LMA and VMA for this section:
                 we state that it should be stored in rom (so that it can
                 actually be programmed onto your microcontroller), but
                 referenced at a location inside our ram segment. */
/* For our final directive, we need to know the location in ROM to load
   the data _from_ at the start of execution. */
__sram_func_loadaddr = LOADADDR(.sram_func);

Now that we have a place to put these functions, we can tell GCC to place them there using a small __attribute__ directive:

void exti1_isr(void) {
    // Body of a time-sensitive interrupt goes here
    gpio_set(GPIOA, GPIO1);
    // [...]
    gpio_clear(GPIOA, GPIO1);

With the addition of our __attribute__ field, gcc now knows to keep our function in a new section, and we can verify this with objdump:

$ arm-none-eabi-nm -f sysv -C my_elf_file
Name                  Value   Class        Type         Size     Line  Section
exti0_isr           |080017fc|   W  |              FUNC|00000014|     |.text
exti1_isr           |20000330|   T  |              FUNC|00000028|     |.sram_func
exti2_isr           |080017fc|   W  |              FUNC|00000014|     |.text

As you can see, the ISR we just tagged as being destined for sram_func is no longer in the .text section, and the location of the symbol is not in the 0x0800 0000 ROM section, but the 0x2000 0000 SRAM area. So far so good, but if we were to deploy this code to the device now, as soon as we actually triggered the ISR we would almost certainly encounter a segmentation fault.

This is because we’ve told the linker that this code is located in RAM, but haven’t actually set up a method to actually move that code into RAM - when the microcontroller resets, that memory space will be initialized to junk. In order to fix this, we need to add some code at the very start of our application to actually read the data for the sram_func section out of ROM and copy it into RAM, where it can then be called. To do so, we use the three variables we defined earlier as part of the linker script:

// The variables defined in our linker script are available to us as externed
// unsigned words, the locations of which denote our sections.
extern unsigned __sram_func_start, __sram_func_end, __sram_func_loadaddr;

// Our sram_func_start and sram_func_end variables are located at the start and
// end of the memory space that we want to copy data into.
// The loadaddr variable is located in ROM, at the start of where the data to
// be copied is stored.
// Using these three variables, we can quickly copy the code across.
volatile unsigned *src = &__sram_func_loadaddr;
volatile unsigned *dest = &__sram_func_start;
while (dest < &__sram_func_end) {
  *dest = *src;

Now that we’ve done that, let’s verify that this is indeed faster. As a testbed, I have a STM32F7 series MCU running at 216MHz. I have configured an EXTI interrupt that is triggered on rising edges, and when fired sets and clears a GPIO. The pin driving the EXTI interrupt is then connected to a function generator running at 1Hz, and the input function and GPIO pin are connected to a scope. Here is a representative trigger of the ISR running from flash memory, where the blue trace is the input signal, and the purple trace is the pin toggled by the ISR:

Interrupt latency running from flash

We can see that after the trigger pin goes high, we have a delay of almost exactly 400ns before the ISR triggers and pulls the GPIO pin high. At 216MHz, that’s close to 100 clock cycles! This is rather poor overhead, and for applications that make extensive use of interrupts, the delays will add up. Now let’s see what happens when we instead run our ISR out of SRAM:

Interrupt latency running from SRAM

Not bad! The latency between the trigger going high and the GPIO going high has been cut in half, and the total execution time of the ISR has also dropped by a little under 200ns itself.


While loading code into memory can be useful for performance, it comes at the obvious tradeoff of taking up additional space. Luckily, on some ARM platforms exists a section of memory called ITCM, or the Instruction Tightly Coupled Memory. Unlike its sister memory, the DTCM, the ITCM block can only be accessed by the CPU, and not at all by peripherals such as DMA controllers. It is also not located in a contiguous memory space with the rest of the system memories: at least on the STM32F750, it is located at address 0x0000 0000, unlike the rest of the volatile memories located at 0x2000 0000. Since it’s so isolated, if you have it it is an excellent location for any functions you may want to load into RAM.

Block diagram of STM32F750 with ITCM highlighted

To use the ITCM as a space for functions that need to be able to execute quickly, we can edit our linker script from above with two small changes:

    rom (rx) : ORIGIN = 0x08000000, LENGTH = 64K
    ram (rwx) : ORIGIN = 0x20000000, LENGTH = 320K
    /* Here we add a new memory region: the ITCM RAM */
    itcm (rwx) : ORIGIN = 0x00000000, LENGTH = 16K

.sram_func : {
    __sram_func_start = .;
    . = ALIGN(4);
    __sram_func_end = .;
} >itcm AT>rom /* Instead of ram, we load to itcm */
__sram_func_loadaddr = LOADADDR(.sram_func);

No changes need to be made to our code that loads the function data from ROM, since after a recompilation the variables we used will automatically point at the new data location. Any functions loaded into memory in this block will not count against the memory space available to the program for heap, stack and globals.


Eagle-eyed readers may also notice another memory hidden away in the block diagram above:

Block diagram of STM32F750 with I-Cache highlighted

Regardless of whether you use the manual memory loading above (which I would recommend for methods you need to guarantee run without wait states), you can gain an often significant general performance enhancement simply by enabling the L1 I-cache built directly into the ARM core (on models that have it, that is. Consult your data sheet!)

Unlike the ITCM RAM, the cache does not require manual management other than being turned on. To do so, we can follow the instructions from the ARM V7-M Architecture Reference Manual in section B2.2, “Caches and branch predictors”. As the document states, all caches are disabled at startup. To enable the I-cache, we need to first invalidate it, and then set a bit in the Cache Control Register to enable it. We only need two registers for this, so the code is relatively straightforward:

// Writing to the ICIALLU completely invalidates the entire instruction cache
#define ICIALLU (*(volatile uint32_t *)(0xE000EF50))
// The Configuration and Control Register contains the control bits for
// enabling/disabling caches
#define ARM_CCR (*(volatile uint32_t *)(0xE000ED14))

// We will also define two macros for data and instruction barriers
define __dsb asm __volatile__("dsb" ::: "memory")
define __isb asm __volatile__("isb" ::: "memory")

// Synchronize
__dsb; __isb;
// Invalidate the instruction cache
// Re-synchronize
__dsb; __isb;
// Enable the I-cache
SCB->CCR |= (1 << 17);
// Force a final resync and clear of the instruction pipeline
__dsb; __isb;

You’ll note the inclusion of several dsb and isb blocks - since we’re messing with instruction caches, it’s a good idea to add some explicit synchronization barriers to the code. We will use DSB to prevent execution from continuing before all in-flight memory accesses are complete. We will also use ISB to flush the processor pipeline, forcing all following instructions to be re-fetched. Note that our inline assembly also specifies the ‘memory’ clobber flag since it may change the global state, and so should not be reordered away by the compiler.

When this cache is enabled, common codepaths will start to benefit passively. Even our already-optimized ISR will gain a little more performance from I-caching of GPIO manipulation code called from the ISR that wasn’t itself loaded into ITCM ram:

Interrupt latency running from RAM with I-Cache enabled

With these two tricks, you can unlock a significant amount of extra performance from your embedded system, and can even go further by enabling the D-cache also present in higher-spec ARM cores. However, be advised that with the D-cache comes pitfalls when it comes to non-TCM memories and DMA transfers.

STM32 & OpenCM3 Part 3: CANBus

Companion code for this post available on Github

This is the fourth post in a series on the STM32 series of MCUs and libopencm3. The previous post, on SPI and DMA, can be found here.

What is CANBus?

The CAN bus is a multi-master, low data rate bus for communicating between controllers in environments with potentially high EMI. Initially designed for automotive applications, it is becoming increasingly used in general automation environments as well as by hobbyists. Electrically, CAN uses a differential pair of signals, CANH and CANL, to send data on the bus. In order to transmit a logic ‘1’ (also known as ‘recessive’ in CAN parlance), the differential voltage of the lines is left at 0. To transmit a logic ‘0’ (dominant), the voltage between the lines is driven high. This means that any node transmitting a 0 will override the transmission of a node that is simultaneously trying to transmit a 1. It is this mechanism that allows for the priority system in a CAN network - since each CAN message begins with the message ID, starting from the MSB, any controller asserting a logic ‘0’ on the bus will clobber a controller attempting to transmit a logic ‘1’. Since all transmitters read the bus as they transmit, this clobbering can be detected by the controller with the lower priority transmission, which will back off until the bus is clear again. This protocol is therefore categorized as ‘CSMA/CD+AMP’, or Carrier Sense Multiple Access / Collision Detection + Arbitration on Message Priority.

CAN Signalling, courtesy of Wikipedia

Why would I use CAN?

When transferring data between two microcontroller systems, people are probably already familiar with I2C and SPI, which are commonly used for low (I2C, 100-400kHz) or high (SPI, 100MHz+) speed data transfer between ICs. However, both of these protocols are really intended for operation over a short distance, ideally on the same board. Running I2C or SPI off-board, even for relatively short distances, can start to result in bit errors at higher speeds or in the presence of interference. The electrical integrity problems with I2C and SPI can be alleviated by using differential signals, as is the case with RS422/485. This allows RS485 to transmit data at high (multiple megabit) speeds over distances of 300-900 feet. This might satisfy our reliability or distance requirements, but none of these protocols bake in support for multi-master communication - SPI is very strongly based around a single-master design, and while I2C does allow for multiple devices to control the bus, there is no built-in arbitration support. Similarly for RS485, the application developer must roll their own packet structure and arbitration to handle bus contention.

CANBus performs quite well on some of these points, being:

However, CANbus does have some drawbacks that make it a poor fit for other applications. These include:

When deciding whether or not to use CAN, be sure to think carefully about the requirements of your application and whether or not CAN is the best fit.

Electrical specifications

For ‘High speed’ CAN (~512 Kbps), all controllers (nodes) in the system must be connected to a linear bus, with appropriate termination. This is to mitigate signal reflections, which can cause bit errors at receiving nodes. This does however mean that CAN buses can be slightly more work to add or remove nodes from, compared to systems that allow a ‘star’ topology (e.g. an ethernet switch). Instead each node must be connected directly to a previous node and to a subsequent node, or, in the case of the last node on either end, a terminating resistor of 120 ohms.

If one is willing to sacrifice some speed, ‘fault tolerant’ CAN (~128 Kbps) can be operated in a star topology, with the termination divided up and placed at each node. For more information, the Wikipedia page on CAN has some diagrams.

As an example implementation, I have created a small demo board in KiCad with switchable termination to be used for high-speed CAN communication. The design files are available here if you are interested in producing some yourself, or you can directly order them from PCBway here.

CAN Demo Board using STM32F091

Message format

CAN frames follow a defined format: all standard frames have an 11-bit identifier and up to 8 bytes of data. Extended frames allow 29 bit identifiers, but only the same 8 bytes of data. CAN frames also include checksums, and most CAN implementations in microcontrollers will automatically insert / verify checksums in hardware. The appearance on the wire of CAN frames is as follows:

CAN frame formats

As may be clear from the 8 byte max payload size, CAN is not a good choice for applications that need to transfer large quantities of data. Instead it is much more suited for controls and small sensor data.

N.B: The ‘RTR’ bit in a CAN message is mutually exclusive with the data segment.

If you set the RTR bit, you may still specify a data length code (DLC) but the peripheral will not transmit any data bytes. Be careful when receiving frames that you ignore any data bytes ‘received’ in RTR frames, as they will simply be junk memory, which can led to pernicious bugs.

Using CAN with libopencm3

Now that we have an understanding of the CAN bus architecture, let’s actually build a small application that will send and receive data on the bus. Setting up the basics is relatively straightforward with a call to can_init():

// Enable clock to the CAN peripheral

// Reset the can peripheral

// Initialize the can peripheral
auto success = can_init(
    CAN1, // The can ID

    // Time Triggered Communication Mode?
    // http://www.datamicro.ru/download/iCC_07_CANNetwork_with_Time_Trig_Communication.  pdf
    false, // No TTCM

    // Automatic bus-off management?
    // When the bus error counter hits 255, the CAN will automatically
    // remove itself from the bus. if ABOM is disabled, it won't
    // reconnect unless told to. If ABOM is enabled, it will recover the
    // bus after the recovery sequence.
    true, // Yes ABOM

    // Automatic wakeup mode?
    // 0: The Sleep mode is left on software request by clearing the SLEEP
    // bit of the CAN_MCR register.
    // 1: The Sleep mode is left automatically by hardware on CAN
    // message detection.
    true, // Wake up on message rx

    // No automatic retransmit?
    // If true, will not automatically attempt to re-transmit messages on
    // error
    false, // Do auto-retry

    // Receive FIFO locked mode?
    // If the FIFO is in locked mode,
    //  once the FIFO is full NEW messages are discarded
    // If the FIFO is NOT in locked mode,
    //  once the FIFO is full OLD messages are discarded
    false, // Discard older messages over newer

    // Transmit FIFO priority?
    // This bit controls the transmission order when several mailboxes are
    // pending at the same time.
    // 0: Priority driven by the identifier of the message
    // 1: Priority driven by the request order (chronologically)
    false, // TX priority based on identifier

    //// Bit timing settings
    //// Assuming 48MHz base clock, 87.5% sample point, 500 kBit/s data rate
    //// http://www.bittiming.can-wiki.info/
    // Resync time quanta jump width
    CAN_BTR_SJW_1TQ, // 16,
    // Time segment 1 time quanta width
    CAN_BTR_TS1_11TQ, // 13,
    // Time segment 2 time quanta width
    CAN_BTR_TS2_4TQ, // 2,
    // Baudrate prescaler

    // Loopback mode
    // If set, CAN can transmit but not receive

    // Silent mode
    // If set, CAN can receive but not transmit

// Enable CAN interrupts for FIFO message pending (FMPIE)

// Route the CAN signal to our selected GPIOs
const uint16_t pins = GPIO11 | GPIO12;
gpio_mode_setup(GPIOA, GPIO_MODE_AF, GPIO_PUPD_NONE, pins);
gpio_set_af(GPIOA, GPIO_AF4, pins);

In order to receive messages, in our CAN ISR we need to check to see which FIFO has pending data, and can then read off the message. For this demo, we’ll just put all of the messages in the same queue to be processed later.

void cec_can_isr(void) {
    // Message pending on FIFO 0?

    // Message pending on FIFO 1?

void receive(uint8_t fifo) {
    // Copy CAN message data into main ram
    Frame frame;
                fifo, // FIFO id
                true, // Automatically release FIFO after rx
                &frame.id, &frame.extended_id, &frame.rtr, &frame.filter_id,
                &frame.len, frame.data, &frame.ts);

    // Push the received frame onto a queue to be handled later


So far, our application will receive and try to store all messages that appear on the bus. But for many applications, we may be able to ignore a lot of messages, and save ourselves some CPU time. To this end, the CAN peripheral on the STM32F091 has a series of filter banks that can be used to selectively accept different message types. The general structure of the filters is that you have an ID register used to input the data you want to match against, and then a mask register that defines which bits of ID register are to be matched. This can be a bit complex at first glance - let’s take a look at the relevant figure in the ST reference manual:

Filter Bank Configuration

As an example, let’s say that we have a device that only wants to receive two types of message:

Since these are both standard frames, we can use 16 bit filters, to save space. From figure 315 we can see that the first 11 bits of the register match against the ID, and bit 4 in the lower byte matches the RTR flag in the CAN message. So for our first filter, we want to assert that the message ID is <= 255. Since 255 is 0xFF, or 8 bits set, we know that any ID numbers above 255 will have one of bits 9-11 set. So to match only lower IDs, we can assert that the top three bits of the ID are zero. So for our first filter, we can create it like so:

const uint16_t id1 = 0;               // We want to assert the high bits are zero
const uint16_t mask1 = (0b111 << 12); // The only bits we want to compare are STDID[10:8]

For our second filter, we want to match the ID exactly, so we will load our ID register with our actual desired message value (342) and in our mask we will select all bits of the STDID field. Since we want to assert that the RTR field is also set, we will likewise place a 1 both the ID and MASK registers at bit 5, like so:

const uint16_t id1 = (
    (342 << 5) | // STDID
    (1 << 4)     // RTR
const uint16_t mask1 = (
    (0b11111111111 << 5) | // Match all 11 bits of STDID
    (1 << 4)               // Match the RTR bit

Once we have our filters, we can configure the CAN peripheral with them like so. All messages that match either of these filters will be placed into FIFO 0.

// Create a filter mask that passes all critical broadcast & command
// CAN messages
    0,          // Filter number
    id1, mask1, // Our first filter
    id2, mask2  // Our second filter
    0,          // FIFO 0
    true);      // Enable

Putting it together

Now that we have our CAN peripheral initialized, let’s write a simple demo application. We’ll use the demo board mentioned above (which you can order directly from PCBWay here) to create a simple program that forwards bytes from the UART over the CAN bus. In our main application loop, we’ll first take any characters that have been received over the UART and transmit them over CAN. (Implementation details of the Frame class can be seen here for those curious.)

  // Loop over any characters pending in the UART Rx buffer,
  // and send each one over the CAN bus as a single message.
  char c;
  while (Uart::get(&c)) {
    // Turn on our activity LED
    gpio_set(GPIOB, GPIO12);
    // Echo this character back to the serial console so we can see what
    // we've typed
    // Create a new CAN Frame holder
    CAN::Frame frame;
    frame.id = 1; // Our message ID
    frame.extended_id = false; // This is not an extended ID
    frame.rtr = false; // This is not a request to transmit
    frame.len = 1; // We intend to send one data byte
    frame.data[0] = c; // Our uart character is the first datum
    CAN::transmit(frame); // Send the frame to the CAN output mailbox
    gpio_clear(GPIOB, GPIO12); // Clear our activity LED

We also need to receive frames off the bus and display the data. The receive interrupt we wrote earlier will queue the frames, so we can pop them off in order and print out the details.

  // Loop over any CAN frames pending in the CAN buffer, and print out
  // the ID of the message and all the data bytes.
  CAN::Frame frame;
  while (CAN::pop(frame)) {
    // Turn on our CAN activity LED
    gpio_set(GPIOB, GPIO13);
    // Print the frame ID and all data bytes as hex and plain characters
    printf("Rx ID: %u Data: ", frame.id);
    for (int i = 0; i < frame.len; i++) {
      printf("%02x:%c", frame.data[i], frame.data[i]);
    // Turn off the activity LED
    gpio_clear(GPIOB, GPIO13);

The full firmware listing can be found here.

In order to test this, we can assemble two test boards and flash the same firmware to each. We can then connect the CANH and CANL pins of each board using jumpers, and configure the termination using jumpers. Since each board is connected to only one other board, we will set the jumper position for the connected header to pins 2-3, which connects the jumper pins directly to the transceiver. For the other set of jumpers, we select pins 1-2 to connect the terminating resistors (in this case a split termination of two 59 Ohm resistors and a 4.7nF capacitor) to the bus.

Two demo boards connected up

Once the boards are connected, we can connect a USB to UART adapter to each one and try sending some data back and forth. If everything is working properly, typing into the console of one board will cause it to send characters over CAN to the other, and vice versa.

Communicating over CANBus

This concludes our overview of CANBus, and the implementation details of the CAN peripheral on the STM32 series of microcontrollers. Using the basics in this post you should be able to create far more interesting applications.

As per usual, the code for this post is available on Github.

Next Page >