Make hacks: embedded version info, detecting non timestamp change and automatically generated dependencies

For the purposes of traceability you may wish to embed the version of a program into it. If you’re using full on CI, you very much shouldn’t need this. If your CI system doesn’t record such things, then you need to fix it now.

But if you’re hacking around locally, especially for research it can be really useful to know where that executable came from or more specifically where some results came from, because it’s easy to lose track of that. An easy way to do this is to embed the git hash of the repository into the executable so it can be written alongside the data.

Essentially if git status --porcelain prints nothing then the repository is entirely clean. With that in mind, here’s a short script which generates a C++ source file with the git hash if the repository is clean, and prints a loud, red warning if it is not clean. Here is get_version.sh:

git_hash=`git rev-parse HEAD`

if [[ "$(git status --porcelain)" != "" ]]
then
	echo -e "\033[31mThere are uncommitted changes. This means that the build" 1>&2
	echo -e "Will not represent a traceable version.\033[0m" 1>&2
	time=`date +%s`
	version="${git_hash}-${time}"
else
	version="${git_hash}"
fi

cat <<FOO
namespace version{
	const char* version_string = "$version";
}
FOO

It’s easy enough to use from make:

.PHONY: FORCE
versioninfo.cc: FORCE
	bash get_version.sh > versioninfo.cc

and of course make your program depend on versioninfo.o. But it’s not very clean; this will rebuild and then re-link every single time. The key is to make the FORCE dependency depend on whether anything has changed.

This script (get_version_target.sh), reruns get_version.sh and compares it to the existing versioninfo.cc. If there’s a change, it prints the target (FORCE), otherwise it prints nothing.

if ! diff -q versioninfo.cc <( bash get_version.sh 2> /dev/null ) > /dev/null 2>&1
then
        echo FORCE
fi

You then need to plumb this into make using the shell command:

version_target=$(shell bash get_version_target.sh)
.PHONY: FORCE 
versioninfo.cc: $(version_target)
	bash get_version.sh > versioninfo.cc

This will now only re-generate versioninfo.cc (and hence the .o and the final executables) if the git hash changes.

With the basics in place, you can make the version info as detailed as you like, for example you could record any tags and branch names, so you could record those and if it’s an actual point release etc. The downside of the shell commands is that they run every time make is called so you will want to make them fast otherwise incremental rebuilds will become annoyingly slow.

Using this mechanism, make can be much more malleable than expected. This is an immensely powerful feature. But remember:

Realtime AR world transformations with occlusions

This is me, my team and collaborators have been working on recently. World transforming AR specifically the floor.

You can see occlusions, such as the pillars occluding the floor effect, but we have more sophisticated occlusion handling too:

You can’t tell from this video that the occlusion handling is dynamic, so if the postbox managed to move, the occlusions would stay up to date. And here’s a gallery of nice shots

If you have Snapchat and want to try it for yourself, here are the snapcodes:

Adafruit mini thermal printer, part 3/?: Long jobs, cancellation and paper out

Writing a printer driver from scratch is quite involved. Who knew?

Code on github: https://github.com/edrosten/adafruit-thermal-printer-driver. Note: I wrote these posts as I went along so there may be bugs in the code snippets which are fixed later. I recommend checking the GitHub source before using a snippet.

This post appears to be about three unrelated things but it isn’t. It’s all about reading back data from the printer.

Cancellation

So, cancellation works in as much as things stop printing. Except none of the end of job stuff gets printed (the “cancelled” message and the paper eject). First I thought it was because I was lazy, so I changed the signal handler to:

	{
		struct sigaction int_action;
		memset(&int_action, 0, sizeof(int_action));
		sigemptyset(&int_action.sa_mask);
		int_action.sa_handler = [](int){
			cancel_job = 1;
		};
		sigaction(SIGTERM, &int_action, nullptr);
	}

This is the approved method, since the signal method is ill specified in general and on Linux on entry to the handler, it causes the handler to get reset to the default (terminate). I thought maybe that was happening. Do you think this worked?

The next step was to add LogLevel debug to /etc/cups/cupsd.conf, so it records all my debug messages. It does, along with a bunch of other useful stuff and its all indexed by the print job number. A filtered log looks like this:

D [29/Dec/2019:14:21:18 +0000] [Job 184] envp[25]=\"PRINTER=pl\"
D [29/Dec/2019:14:21:18 +0000] [Job 184] envp[26]=\"PRINTER_STATE_REASONS=none\"
D [29/Dec/2019:14:21:18 +0000] [Job 184] envp[27]=\"CUPS_FILETYPE=document\"
D [29/Dec/2019:14:21:18 +0000] [Job 184] envp[28]=\"FINAL_CONTENT_TYPE=application/vnd.cups-raster\"
D [29/Dec/2019:14:21:18 +0000] [Job 184] envp[29]=\"AUTH_INFO_REQUIRED=none\"
D [29/Dec/2019:14:21:19 +0000] [Job 184] Start rendering...
D [29/Dec/2019:14:21:19 +0000] [Job 184] Set job-printer-state-message to "Start rendering...", current level=INFO
D [29/Dec/2019:14:21:19 +0000] [Job 184] Processing page 1...
D [29/Dec/2019:14:21:19 +0000] [Job 184] Set job-printer-state-message to "Processing page 1...", current level=INFO
D [29/Dec/2019:14:21:19 +0000] [Job 184] PAGE: DEBUG: Read 2 bytes of print data...
D [29/Dec/2019:14:21:19 +0000] [Job 184] 1 1
D [29/Dec/2019:14:21:19 +0000] [Job 184] bitsperpixel 8
D [29/Dec/2019:14:21:19 +0000] [Job 184] BitsPerColor 8
D [29/Dec/2019:14:21:19 +0000] [Job 184] Width 384
D [29/Dec/2019:14:21:19 +0000] [Job 184] Height799
D [29/Dec/2019:14:21:19 +0000] [Job 184] feed_between_pages_mm 0
D [29/Dec/2019:14:21:19 +0000] [Job 184] mark_page_boundary 0
D [29/Dec/2019:14:21:19 +0000] [Job 184] eject_after_print_mm 10
D [29/Dec/2019:14:21:19 +0000] [Job 184] auto_crop 0
D [29/Dec/2019:14:21:19 +0000] [Job 184] enhance_resolution DEBUG: Wrote 2 bytes of print data...
D [29/Dec/2019:14:21:19 +0000] [Job 184] 0
D [29/Dec/2019:14:21:19 +0000] [Job 184] Feeding 155 lines
D [29/Dec/2019:14:21:19 +0000] [Job 184] Feeding 47 lines

It has the outputs from various filters all mixed together, possibly with some race conditions… (can you spot them?). Anyway, the cancel message is coming through and getting processed correctly. But no output is happening.

Debugging this was tricky because there were several causes. What I eventually did was add a 100ms pause between lines in order to reduce the amount of paper wasted and that revealed something interesting. One case was simply that sometimes the heating level was too low and the text was invisible.

In the other case, I’m just not sure. If the buffer is too full, then the last bits of the job seem to get “lost” somehow, if a cancellation occurs. With a 100ms pause, I always get the cancellation message. If I make the pause shorter then the printer can’t keep up and after a while the buffers all become full. In that case, I get cancellation messages if done early (when the buffers aren’t yet full) but not late.

I don’t yet know if long jobs get truncated. I suspect that the same would happen because there appears to be nothing functionally different between cancellation and normal termination. I don’t know who is responsible for this, but I’d be surprised if it was CUPS. My guess is no one has ever tested printing large amounts of full page bitmaps on this printer simply because that’s not the intended use. Speaking of not the intended use…

Abuse of paper sensors

As far as I can tell there isn’t an obvious way to query the buffer status to avoid it getting too full. I don’t even know where the buffer is. I expect the USB system has one, as does the USB chip and the UART on the printer.

But the printer does have a “Transmit Status” command (Page 42) for which it warns that there may be a lag since it’s processed in sequence. Even worse/better you can’t use this one to detect paper out because once the paper ends, the printer goes offline and won’t execute the command (I expect the paper sensor status command may be more asynchronous). Also that appears to be untrue, I tried it with the following code:

exec 3<> /dev/usb/lp0
echo -ne '\x1dr1' >&3
dd bs=1 count=1 status=none <&3 | od -td1 

And I got back 0 with paper in and 12 with the door open.

That apparently useless synchronous mechanism may be just the ticket: I bet if I stuff the command stream with these then I can get an approximation of the number of lines printed. The code looks something like this:

void transmit_status(){
	cout << GS << "r1" << flush;
}


void wait_for_lines(const int lines_sent, int& read_back, int max_diff){
	for(;;){
		char buf;
		ssize_t bytes_read = cupsBackChannelRead(&buf, 1, 0.0);

		if(bytes_read > 0)
			read_back++;

		if(lines_sent - read_back <= max_diff)
			break;

		cerr << "DEBUG: buffer too full (" << lines_sent - read_back << "), pausing...\n";
		using namespace std::literals;
		std::this_thread::sleep_for(100ms);
	}
}

// ... and in the main print loop...
			//Stuff requests for paper status into the command stream
			//and count the returns. We allow a gap of 80 lines (1cm of printing)
			transmit_status();
			lines_sent++;
			wait_for_lines(lines_sent, read_back, 80);

Checking the print logs shows this does what is expected. Furthermore, cancellation works properly (it prints the cancelled message and ejects the job) and is pretty quick!

Paper out!

OK, so I’m already reading the paper status. The manual suggests I might not be as I mentioned except I’m reading it before/after every line, so in that case I think I’m safe. Besides, it’s not entirely clear how you’re meant to differentiate between all the async replies:

When Auto Status Back (ASB) is enabled using GS a, the status
transmitted by GS r and the ASB status must be differentiated using.

(page 42)

Maybe some of the undefined bits are actually set. Who knows?

Anyway, all that remains is to transmit that back to CUPS. It’s broadly covered here.

BAH!

It didn’t work. Turns out the manual is only not right in very specific circumstances. Fortunately it seems for the status command bit 5 is always set so I could test for that.

So I stuffed the command stream with the proper status reports too and, well, guess what?

I just got a big old stream of zeros back from the printer. I could try the async reporting. That might work, but the printer has only a single sensor and stops running when it’s tripped. What I could do is see if nothing has changed for some time and report that as a paper out event.

This seems a bit hacky and it is. I’m not all that surprised though. This family of printers are mostly RS/232 based with asynchronous status lines in addition for paper, not USB. They’re also not expected to print out vast amounts of data; receipts are usually a few pages at most of plain text. I expect these obscure paths haven’t been exercised much.

Oh yes, hacky. So, here’s the code, it’s pretty straightforward overall:

void wait_for_lines(const int lines_sent, int& read_back, int max_diff){
	using namespace std::literals;
	using namespace std::chrono;

	auto time_of_last_change = steady_clock::now();
	bool has_paper=true;

	for(;;){
		char buf;
		ssize_t bytes_read = cupsBackChannelRead(&buf, 1, 0.0);

		if(bytes_read > 0){
			read_back++;
			
			if(!has_paper){
				cerr << "STATE: -media-empty\n";
				cerr << "STATE: -media-needed\n";
				cerr << "STATE: -cover-open\n";
				cerr << "INFO: Printing\n";
			}
				
			has_paper = true;
			time_of_last_change = steady_clock::now();
		}
		else if(auto interval = steady_clock::now() - time_of_last_change; interval > 2500ms){
			cerr << "DEBUG: no change for " << duration_cast<seconds>(interval).count() << " seconds, assuming no paper\n";
			if(has_paper){
				cerr << "STATE: +media-empty\n";
				cerr << "STATE: +media-needed\n";
				cerr << "STATE: +cover-open\n";
				cerr << "INFO: Printer door open or no paper left\n";
			}
			has_paper = false;
		}

		cerr << "DEBUG: Lines sent=" << lines_sent << " lines printed=" << read_back << "\n";

		if(lines_sent - read_back <= max_diff)
			break;

		cerr << "DEBUG: buffer too full (" << lines_sent - read_back << "), pausing...\n";
		std::this_thread::sleep_for(100ms);
	}
}

I’ve gone for an all inclusive approach with the messages. The printer cannot distinguish between the door being open and a lack of paper, so I’ve reported both.

It works!

The driver is now feature complete for a first version at any rate. There’s some minor image quality problems in normal mode (caused by fast feeds before bitmaps) and a bit of stripyness caused by poor calibration in enhanced mode. And the plain text filter probably should be a proper filter that does status read back and buffering. But it isn’t.

Adafruit mini thermal printer, part 2/?: CUPS and other vessels

I bought a printer and have blogged about it because it’s literally the most interesting thing ever.

Code on github: https://github.com/edrosten/adafruit-thermal-printer-driver. Note: I wrote these posts as I went along so there may be bugs in the code snippets which are fixed later. I recommend checking the GitHub source before using a snippet.

This post is about integrating with CUPS so I can print from normal programs.

Integrating with CUPS

So, I have a sort of working example of CUPS integration in the existing ZJ-58 driver. It is, I suspect not very good. Nonetheless, I’ll start there since it’s vastly easier starting from a working example than the documentation.

Note from the future: The documentation…

It does exist, but it’s scattered over the various projects, those being PostScript, other Adobe printer guff, CUPS, the GhostScript interpreter, the printer working group and so on). It’s the type of documentation where you can’t find anything so you do 95% of the work the hard way, get stuck on an obscure API call/keyword/etc and then that string turns up the documentation you needed at the beginning.

Anyway, Here’s the install script from the existing driver:

#!/bin/bash

# Installs zj-58 driver
# Tested as working under Ubuntu 14.04

/etc/init.d/cups stop
cp rastertozj /usr/lib/cups/filter/
mkdir -p /usr/share/cups/model/zjiang
cp ZJ-58.ppd /usr/share/cups/model/zjiang/
cd /usr/lib/cups/filter
chmod 755 rastertozj
chown root:root rastertozj
cd -
/etc/init.d/cups start

That’s pretty simple: basically it dumps some files into the CUPS tree and restarts CUPS.

From what I understand, CUPS essentially has some sort of specification of various filter chains (which can vary based on the input, e.g. a plain text file, a postscript file and a JPG will have different input filters). A given filter lists its accepted inputs and CUPS works backwards to figure out how to generate what’s required. For raster things (i.e. not plain text when the printer can accept plain text) CUPS will rasterise the input and you need to then get it sent to the converter to convert it to the right control codes (mostly the topic of the previous post).

Many of those are controlled by a PPD file. This stands for “PostScript Printer Description” and tells CUPS all about the printer capabilities. It also has extensions beyond the Adobe PPD spec to allow you to specify rasterisation and filters for non PostScript printers.

The driver of course comes with a PPD (it has to), but it’s long, complicated, has fragments of PostScript in it and doesn’t even pass the tests run by the cupstestppd command. And there’s a lot of duplicated information about pages sizes which I suspect needs to be consistent. Not great but it’s a start.

So, while PPD is documented (or some approximation thereof) and the CUPS extensions are likewise, apparently you’re not really meant to write PPDs anyway. The easy/approved way is to write DRV files and then compile them into one or more PPD files using ppdc.

Either way the documentation is poor. There are lots of attributes in existing PPD and DRV files like “Filesystem” that it’s very hard to find any kind of documentation for and others like “PSVersion” which are weakly documented (what are the acceptable range of values?).

Some information is here. On the subject of PSVersion, typing revision = into a GhostScript interpreter reveals that my machine (Ubuntu 18.04) has revision 926 for whatever that’s worth. Either way it seems optional. Anyway, I’ve tried to pare down my DRV file to the absolute minimum which covers what I want and I got this:

#include <font.defs>

DriverType custom  //Required I believe to set downstream filters 
ManualCopies Yes //Set to yes if the driver doesn't know how to print multiples of pages
Attribute "LanguageLevel" "" "3" //Default is 2 (from 1991), latest version is from 1997
Attribute "DefaultColorSpace" "" "Gray" //Self explanatory except does this mean something else can change it?
Attribute "TTRasterizer" "" "Type42" //Default is none, Type42 is the only extant useful one.
Filter application/vnd.cups-raster 0 rastertoadafruitmini //Arguments are datatype to feed to the filter, the expected CPU load, and the name of the filter executable
ColorDevice False

Font * //Include all fonts

// Manufacturer, model name, and version of the driver
Manufacturer "Adafruit"
ModelName "Mini"
Version 1.0
ModelNumber 579 //That's the product number on the website.

//I believe this allows users to specify custom sizes in the
//print dialog, or on the command line.
VariablePaperSize Yes
MinSize 58mm 5mm
MaxSize 58mm 1000mm

//#media creates media definitions which may or may not be used
//The paper is always 58mm wide, and have for now three different
//lengths
#media "58x50mm" 58mm 50mm
#media "58x100mm" 58mm 100mm
#media "58x200mm" 58mm 200mm

//The print area is always 48mm wide, centred
HWMargins 5mm 0 5mm 0

//This actually uses the media definitions above
*MediaSize "58x50mm"
MediaSize "58x100mm"
MediaSize "58x200mm"

// Supported resolutions
// Use as: Resolution colorspace bits-per-color row-count row-feed row-step name
// Apparently mostly the row stuff is 0 in most drivers. The last field
// (name) needs to be formatted correctly
*Resolution k 8 0 0 0 "208dpi/208 DPI"

// Name of the PPD file to be generated
PCFileName "mini.ppd"

OK, strictly speaking this isn’t the absolute minimum, since I’ve specified several virtual page sizes and variable sized pages, which is how CUPS deals with roll media. Here’s the corresponding install shell script to dump things in the right place:

/etc/init.d/cups stop

mkdir -p /usr/share/cups/model/adafruit
install rastertoadafruitmini  /usr/lib/cups/filter/rastertoadafruitmini
install ppd/mini.ppd /usr/share/cups/model/adafruit/mini.ppd

/etc/init.d/cups start

Now, running that and going to http://localhost:631 and going through the motions shows the printer there with the options I’d expect (i.e. paper size). The printer device appears as “unknown” in CUPS since it works as a USB parallel port (/dev/usb/lp0), but doesn’t report anything back to CUPS. Even with that , it won’t work yet, because I need in no particular order

  • Proper information logging to stderr in a format that CUPS likes
  • Deal with commandline arguments that CUPS hands me
  • Handle SIGTERM (used to cancel jobs) and not leave the printer in a bad state

In addition, you can add arbitrary choices to the driver which get passed on to the filter so I think I’ll add ones for feeding paper after the job has done (so the end of the last page ends at tearoff on the printer), auto cropping pages (removing white space at the top and bottom–useful for roll media), and marking page boundaries. Because why not? I only have to implement them later.

Options are implemented using an option directive followed by a bunch of choice directives, e.g.

Option "TestOption" PickOne DocumentSetup 0
  *Choice "A" ""
  Choice "B" ""
  Choice "C" ""

You can have Boolean, PickOne or PickMany. I don’t really see the point of Boolean: all of them need to have choice directives (for reasons which will soon become clear), so there’s little difference between a Boolean and a PickOne with two options.

The only difference seems to be that it renders a boolean as a radio group not a drop down list in the web interface:

hmmm. I wonder…

OK Confirmed! You can have as many “boolean” choices as you like, though note that the troolean choices don’t appear in the print dialog boxes, whereas booleans appear as checkboxes. Neither the compiler nor the validator complained which seems like a mild oversight.

With that silly aside out of the way, the next bit is how those options are passed to the printer driver. It turns out there are two ways, both of which are applied simultaneously.

The first, is that the options are passed as a command line argument to the filter, along with the PPD file (in the PPD environment variable). The CUPS API provides some handy functions for parsing PPD files and option strings and generally dealing with it.

The second is that each choice comes with an arbitrary snippet if PostScript code which is run at the point specified by the option directive (it can be at places like document start, page start). Now PostScript has a setpagedevice command which basically accumulates a dictionary for device specific use. The CUPS driver will put certain elements in that dictionary into the raster page headers, and you can access them from C in the filter. It doesn’t support arbitrary dictionaries, and in fact what it has is:

unsigned cupsInteger[16];
float cupsReal[16];
char cupsString[16][64];

You can fill these up by putting appropriately named things into the dictionary, e.g.:

<</cupsInteger1 10 /cupsReal7 2.2 /cupsString3 (a string)>> setpagedevice

W00t! I just found the documentation (by searching for cupsInteger0 to see if it was 0-based or 1-based; it’s 0-based). Turns out there are loads of parameters you can pass this way. Many have “accepted” meanings but you can abuse them to pass arbitrary data since you control both sides.

The two choices are pretty much equivalent, so I’ll pick… uh. Ummm OK wow I’m suffering from choice indecision here. OK, I’ll go for option 2. The API for option 1 is the usual annoying C faff, plus apparently it’s been deprecated since 2012 and I don’t have a nice example of the new API to copy from.

Putting all that together code added to the DRV file looks like this:

//The last argument is the order in which the order in which the options 
//are executed (each one comes with a snippet of code to execute). In this
//case, all snippets are empty.
Option "PageFeed/Feed paper between pages" PickOne DocumentSetup 0
  *Choice "None" "<</cupsInteger0  0>>setpagedevice"
  Choice "1mm"   "<</cupsInteger0  1>>setpagedevice"
  Choice "2mm"   "<</cupsInteger0  2>>setpagedevice"
  Choice "5mm"   "<</cupsInteger0  5>>setpagedevice"
  Choice "10mm" "<</cupsInteger0 10>>setpagedevice"

Option "PageMark/Mark where to cut pages" Boolean DocumentSetup 1
  *Choice "No" "<</cupsInteger1 0>>setpagedevice"
  Choice "Yes" "<</cupsInteger1 1>>setpagedevice"

Option "EjectFeed/Feed paper after printing" PickOne DocumentSetup 2
  Choice "None"  "<</cupsInteger2  0>>setpagedevice"
  *Choice "5mm"  "<</cupsInteger2  5>>setpagedevice"
  Choice "10mm" "<</cupsInteger2 10>>setpagedevice"
  
Option "AutoCrop/Crop page to printed area" Boolean DocumentSetup 3
  *Choice "No" "<</cupsInteger3 0>>setpagedevice"
  Choice "Yes" "<</cupsInteger3 1>>setpagedevice"

The *’s indicate the default choices. And this so far appears to work! The web interface shows this:

And the print dialog in Firefox looks like this:

Sweet!

Writing a valid CUPS filter

This is actually documented reasonably well if you know where to look. I believe I can ignore all arguments (I’m using the other method for options, and I’ve told the driver I don’t know how to make copies myself) except the optional argv[6] which is the file to print if it’s not stdin. Yay.

Cancellation is easy: ignore SIGPIPE and clean up on SIGTERM. Since it’s a simple program, I can use a simple solution where I just poll a global variable:

volatile sig_atomic_t cancel_job = 0;
//...
	signal(SIGPIPE, SIG_IGN);
	signal(SIGTERM, [](int){ cancel_job = 1;});

Logging likewise is easy and involves writing to stderr something like TYPE: data where TYPE is the message type. The type has things such as ERROR, DEBUG, etc for logging, PAGE for recording the current page number, STATE for indicating things like paper empty and so on. The format of the data depends on the message type.

Paper empty and so on can be queried from the printer using special control codes and CUPS looks like it has a way to read back anything returned. I’m not so sure how this works yet. I’ll deal with that later.

Dealing with options took me far too long. I started with the following code snippet:

	cups_raster_t *ras;
	cups_page_header2_t header;
	//...
	while (cupsRasterReadHeader2(ras, &header))
	{
		feed_between_pages_mm = header.cupsInteger[0];
		mark_page_boundary = header.cupsInteger[1];
		eject_after_print_mm = header.cupsInteger[2];
		auto_crop = header.cupsInteger[3];
		enhance_resolution = header.cupsInteger[4];

and it didn’t really work. And by “didn’t work”, I mean that I tried adding -dcupsInteger0=1 to the GhostScript invocation (this sets an integer variable and somehow these magically wind up in setpagedevice, I don’t know how) and I could only set 0, 1 and 2. None of the other integers could be set.

If you cast your mind back to the first post in this series, I mentioned that I cargo-culted an invocation of GhostScript and wasn’t sure what everything did. Well, it came to bite me here. It has the innocuous looking argument -sMediaClass=PwgRaster (-s just sets a variable in the interpreter). This is now getting in quite deep. MediaClass is a variable which affects the setpagedevice command (page 21 of the PostScript® Language Reference Manual Supplement published in 1996 on April 1 and it is deadly serious) in various nonspecific (vendor defined) ways. And one such vendor is the shadowy cabal known as the “Printer Working Group” or PWG for short (its more exciting if they are a shadowy cabal). I sort of unearthed them by forlornly digging through cups/raster.h looking for clues and found this (edited) for display:

// The following PWG 5102.4 definitions specify indices into the
// cupsInteger[] array in the raster header.
#  define CUPS_RASTER_PWG_TotalPageCount	0
#  define CUPS_RASTER_PWG_CrossFeedTransform	1
// etc...
#  define CUPS_RASTER_PWG_VendorLength		15

Turns out they have defined their own meanings for the user-defined extensions and brazenly took all of them. What I don’t understand is why I could set 0, 1 and 2, but not 3 onwards. No clues there. It also stopped cupsReal and cupsString from working and set PWG_AlternatePrimary to 224-1. ¯\_(ツ)_/¯

What went wrong

So that all sort of worked, and I can print out cats using lpr. Except…

Inverted cats. And junk

The cats come out inverted, like this:

meow!

This is because I had:

*Resolution k 8 0 0 0 "203dpi/203 DPI"

which is the “black” colour model. If I change the “k” to “w”, I get what I expect except with some junk at the top.

What I actually need is:

*ColorModel Gray/Grayscale w chunky 0
*Resolution - 8 0 0 0 "203dpi/203 DPI"

I don’t know why. The colour model specifies the white model (along with chunky which is means packed for colour data and no compression), then the resolution says to not modify the colour model. Ok, sure…

Nope!!

Turns out that wasn’t it. I must have just reset things when making that change. The junk was because… well I don’t know exactly. It doesn’t appear on the first printout, it only appears on the third. And if I send enough text to the printer then the next image is fine. It therefore appears as if something was getting flushed before the last line was complete. Then the first few bytes (including the start bitmap control code) were getting eaten up finishing the previous line and then it was printing data out as text.

Turns out the offending bit was this function

void printerInitialise(){
	cout << ESC << '\x40';
}

calls to which I sprinkled liberally around, and these are messing things up. Here’s the funny thing though: putting a cout << flush after the first one fixed it. That ought to make sense: the printer gets data asynchronously then starts processing it while the UART asynchronously fills the receive buffer. It processes the initialise command and loses the first few control codes. Or something.

Except… the symptoms only manifested after several images, making it look like it was state being carried over. It’s weird, I don’t get it. Clearly there’s some internal state somewhere, and part of me things is might be in CUPS because I suspect the original driver used to work just fine.

Page Sizes

The print dialog boxes seemed to get deeply confused about the smallest page size (58x50mm). The reason for this it turns out is that it’s really a landscape page not a portrait one and pages need to be specified in portrait orientation. Except that would make the width wrong. If I’d paid attention to the warnings from cupstestppd, then I would not have had this problem.

ppd/mini.ppd: PASS
        WARN    Size "58x50mm" should be the Adobe standard name "50x58mmRotated".

And it turns out all I have to do is switch the name:

#media "50x58mmRotated" 58mm 50mm
#media "58x100mm" 58mm 100mm
#media "58x200mm" 58mm 200mm

HWMargins 5mm 0 5mm 0

*MediaSize "50x58mmRotated"
MediaSize "58x100mm"
MediaSize "58x200mm"

and things seem to be much more sensible.

Booleans

The print dialog box renderers don’t really know which option is meant to correspond to a check mark and which isn’t. I tried changing the keyword to “True” and “False” and putting true first in the list, e.g.:

Option "PageMark/Mark where to cut pages" Boolean DocumentSetup 1
  Choice "True/Yes" "<</cupsInteger1 1>>setpagedevice"
  *Choice "False/No" "<</cupsInteger1 0>>setpagedevice"

That seemed to do the job. I believe it’s the ordering that matters, I’m not sure though.

Other stuff

There were a few other miscellaneous bits and bobs to fix too. In addition I implemented the various features I mentioned above. I decided also to emit blank lines as a feed rather than a blank line because it’s a fair bit faster. Except I had to suppress that in enhanced resolution mode, because otherwise the first few lines printed after a gap were too dark.

I also want the printer to print plain text as plain text. This isn’t necessary but it’s always been idiomatic to pass through like that, rather than relying on the postscript rasteriser. I can fix that with one extra line in the DRV file:

Filter text/plain 0 -

That tells CUPS that it accepts text, is no cost and to use a null filter program.

Cancellation

Oh wow this turned out to be hard. Way harder than expected because it reveals deep problems. It’s going to be a whole other blog post.

Result!

OK so basically it works!

I can print using lp (or lpr), and set options like -o Enhance=True -o PageMark=True and it obeys them.

Recognise this?

Adafruit mini thermal printer, part 1/?: getting better pictures

I bought an AdaFruit Mini thermal printer.

Code on github: https://github.com/edrosten/adafruit-thermal-printer-driver. Note: I wrote these posts as I went along so there may be bugs in the code snippets which are fixed later. I recommend checking the GitHub source before using a snippet.

It’s pretty cute, and it’s actually very old school in terms of its function. Firstly in a very old fashioned twist, it comes with a full manual documenting every single control code. Not only that but the printer is surprisingly capable and it’s designed to work with very low end driving systems. It doesn’t just print bitmaps, it has various fonts and modes (double height, width, etc), you can download custom fonts and bitmaps to print on demand.  You can print upside down and back to front so the text looks the right way round if you’re facing the printer (super cute!). I has justification modes, bold and underline. It can even print barcodes!

You know this reminds me of when I was 14(?) and got my first computer, a BBC Micro complete with a 5.25″ floppy drive and a printer. The printer came with a manual with full documentation of all the control codes and I devoured them and wrote a basic typesetter like system in which I did my school projects.

cdst5nkwyaew1fw

So where was I?

Oh yes, well I don’t actually need most of those features. I’m planing on driving it from Linux (on a Pi), which means it’ll be driven by GhostScript via CUPS and will print bitmaps. And not use any of those features.

Turns out Adafruit provide a CUPS driver. Apparently provided from one provided by the printer maufacturer? So, I installed it and this is the result:

20191202_181147

woo! it prints! Except… the output isn’t great. The printer is monochrome and the pictures come out halftoned using a halftone screen. While that’s a fine choice for various kinds of printing, it’s not great for a device with independent pixels. For that, a dithering method such as Floyd-Steinberg would be much better. Also it’s messing up the first line and printing junk, but you know details, details.

PostScript, being designed for proper printing has native support for halftoning. It doesn’t for dithering, and it turns out there’s no way to persuade it to emit a monochrome bitmap using dithering instead of halftone screens. If you want dithering, you need to do it in the driver. So, I’m going to need a custom driver.

So I first need to understand printing.

Printing on Linux greatly simplified

Printing on Linux isn’t simple. Partly this is because printing in general is not simple. And partly it’s because printing has changed a lot over the years and there are lots of vestigial bits lying around. For common, modern systems the order of operations is roughly:

  1. CUPS accepts jobs (and provides information to the print dialogs).
  2. CUPS examines the file type and decides what to do next, e.g. whether to run it through GhostScript.
  3. CUPS runs it through ghostscript generating a stream in CUPS raster format. This is a simple bitmap format with a C API.
  4. CUPS runs some arbitrary filter program.
  5. Filter program transforms CUPS bitmap into printer control codes.
  6. CUPS routes the resulting data to the correct device.

GhostScript also has some printer drivers built in, an there are various other filter schemes (GhostScript is one of many) such as foomatic, and of course printers can accept plain text too. I’m not really interested in those so I’ll stick to the sequence above.

Steps 2-4 are controlled by a PPD (PostScript Printer Description) file, and 5 is a program which reads in CUPS bitmap data and emits control codes. The CUPS raster format is well documented but it seems simpler to use the C API, especially as I have a working driver to cadge from.

What I’m going to do first is figure out how to print out what I want (i.e. the right control codes) and then figure out how to work it into CUPS.

Getting CUPS raster data and a simple driver

The first job is to get the input data. After a bunch of cargo-culting, I got this script:

DPI=203.2
gs -dPARANOIDSAFER -dNOPAUSE -dBATCH -sstdout=%stderr -sOutputFile=%stdout \
-sDEVICE=cups -sMediaClass=PwgRaster -sOutputType=Automatic -r${DPI}x${DPI} \
-dDEVICEWIDTH=384 -dDEVICEHEIGHT=384 -dcupsBitsPerColor=8 -dcupsColorOrder=0 \
-dcupsColorSpace=0 -dcupsBorderlessScalingFactor=0.0000 -dcupsInteger1=1 \
-dcupsInteger2=1 -scupsPageSizeName=na_letter_8.5x11in -I/usr/share/cups/fonts \
"$@"


I don’t remember precisely how I found all the various bits. The important things are that it’s colourspace 0 (white), 8 bits per colour, CUPS raster format, 384 pixels wide and 8 pixels per mm. Everything else is just necessary guff (IO redirection, batch and no pause) or irrelevant stuff I never deleted.

I then basically deleted everything except the stream processing from the driver, then deleted that and started writing from scratch. After lots of head scratching and making a lot of mistakes I read the manual more carefully (bitmaps are always a multiple of 8 pixels wide) and got this code up and running:

#include <cups/raster.h>

#include <iostream>
#include <vector>
#include <array>
#include <utility>
#include <cmath>

using std::clog;
using std::cout;
using std::endl;
using std::vector;
using std::array;


constexpr unsigned char ESC = 0x1b;
constexpr unsigned char GS = 0x1d;

// Write out a std::array of bytes as bytes.  This will form the basis
// of sending data to the printer.
template<size_t N>
std::ostream& operator<<(std::ostream& out, const array<unsigned char, N>& a){
	out.write(reinterpret_cast<const char*>(a.data()), a.size());
	return out;
}

array<unsigned char, 2> binary(uint16_t n){
	return {{static_cast<unsigned char>(n&0xff), static_cast<unsigned char>(n >> 8)}};
}


void printerInitialise(){
	cout << ESC << '\x40';
}

// enter raster mode and set up x and y dimensions
void rasterheader(uint16_t xsize, uint16_t ysize)
{
	// Page 33 of the manual
	// The x size is the number of bytes per row, so the number of pixels
	// is always a multiple of 8
	cout << GS << 'v' << '0' << '\0' << binary((xsize+7)/8) << binary(ysize);
}


int main(){

	cups_raster_t *ras = cupsRasterOpen(0, CUPS_RASTER_READ);
	cups_page_header2_t header;
	int page = 0;

	while (cupsRasterReadHeader2(ras, &header))
	{
		/* setup this page */
		page ++;
		clog << "PAGE: " << page << " " << header.NumCopies << "\n";
		clog << "BPP: " << header.cupsBitsPerPixel << endl;
		clog << "BitsPerColor: " << header.cupsBitsPerColor << endl;
		clog << "Width: " << header.cupsWidth << endl;
		clog << "Height: " << header.cupsHeight << endl;

		// Input data buffer for one line
		vector<unsigned char> buffer(header.cupsBytesPerLine);
		
		clog << "Line bytes: " << buffer.size() << endl;
		printerInitialise();

		/* read raster data */
		for (unsigned int y = 0; y < header.cupsHeight; y ++)
		{
			if (cupsRasterReadPixels(ras, buffer.data(), header.cupsBytesPerLine) == 0)
				break;

			//Print in MSB format, one line at a time
			rasterheader(header.cupsWidth, 1);
			unsigned char current=0;
			int bits=0;

			for(const auto& pixel: buffer){
				current |= (pixel>128)<<(7-bits);
				bits++;
				if(bits == 8){
					cout << current;
					bits = 0;
					current = 0;
				}
			}
			if(bits)
				cout << current;
		}

		/* finish this page */
	}
	cout << "\n\n\n";
	cupsRasterClose(ras);
}

To run the program, make an eps file, ideally with a cat in it. Then assuming the above script is called “to_cups.sh” and the compiled executable is called “rastertoadafruitmini”, you can run it with:

bash to_cups.sh cat.eps | ./rastertoadafruitmini | sudo dd of=/dev/usb/lp0 

Note that the quantisation is simply “greater than 128”, and the result is:

Note the false start at the top, and the slightly stretched image due to me converting to EPS badly. The underlying image is this:

CC BY 2.5, Copyright of (c) David Corby

It works! You’ll note I got the colours inverted, because I had 1 for white, and 0 for black, whereas 1 means print a pixel (i.e. black). The black bar is because of that and the page being white. The funny thing is that black areas feel incredibly wasteful of ink even though that makes no sense on a thermal printer.

Dithering the output

Clearly a simple threshold is not a very good way of converting greyscale to black and white. In fact it’s somewhat worse than the original halftoned image. The key is to employ some sort of dithering and this is best done by some sort of error diffusion algorithm.

The process works like this. While going in raster scan order:

  1. Quantize the pixel current to 0 or 255
  2. Work out the error between the quantized output and the pixel
  3. Add fractions of the error to nearby pixels which haven’t been processed yet (this is the error diffusion step)

There are quite a few articles on it, such as this excellent one. The most common/well know algorithm for images is the Floyd-Steinberg dithering algorithm. It’s popular because it’s low resource and efficient on simple processors. Since the target machine for this will be lavishly resourced (a Raspberry Pi) I decided to go for the Jarvis, Judice, Ninke algorithm which is essentially identical to Floyd-Steinberg but with a larger error diffusion window and is more expensive and gives slightly better results.

Here’s the code (with the new bits highlighted):

#include <cups/raster.h>

#include <iostream>
#include <vector>
#include <array>
#include <utility>
#include <cmath>
#include <algorithm>

using std::clog;
using std::cout;
using std::endl;
using std::vector;
using std::array;


constexpr unsigned char ESC = 0x1b;
constexpr unsigned char GS = 0x1d;

// Write out a std::array of bytes as bytes.  This will form the basis
// of sending data to the printer.
template<size_t N>
std::ostream& operator<<(std::ostream& out, const array<unsigned char, N>& a){
	out.write(reinterpret_cast<const char*>(a.data()), a.size());
	return out;
}

array<unsigned char, 2> binary(uint16_t n){
	return {{static_cast<unsigned char>(n&0xff), static_cast<unsigned char>(n >> 8)}};
}


void printerInitialise(){
	cout << ESC << '\x40';
}

// enter raster mode and set up x and y dimensions
void rasterheader(uint16_t xsize, uint16_t ysize)
{
	// Page 33 of the manual
	// The x size is the number of bytes per row, so the number of pixels
	// is always a multiple of 8
	cout << GS << 'v' << '0' << '\0' << binary((xsize+7)/8) << binary(ysize);
}


constexpr array<array<int, 5>, 3> diffusion_coefficients = {{
		{{0, 0, 0, 7, 5}},
		{{3, 5, 7, 5, 3}},
		{{1, 3, 5, 3, 1}}
}};
constexpr double diffusion_divisor=42;


int main(){

	cups_raster_t *ras = cupsRasterOpen(0, CUPS_RASTER_READ);
	cups_page_header2_t header;
	int page = 0;

	while (cupsRasterReadHeader2(ras, &header))
	{
		/* setup this page */
		page ++;
		clog << "PAGE: " << page << " " << header.NumCopies << "\n";
		clog << "BPP: " << header.cupsBitsPerPixel << endl;
		clog << "BitsPerColor: " << header.cupsBitsPerColor << endl;
		clog << "Width: " << header.cupsWidth << endl;
		clog << "Height: " << header.cupsHeight << endl;

		// Input data buffer for one line
		vector<unsigned char> buffer(header.cupsBytesPerLine);
		
		//Error diffusion data
		vector<vector<double>> errors(diffusion_coefficients.size(), vector<double>(buffer.size(), 0.0));

		clog << "Line bytes: " << buffer.size() << endl;
		printerInitialise();

		/* read raster data */
		for (unsigned int y = 0; y < header.cupsHeight; y ++)
		{
			if (cupsRasterReadPixels(ras, buffer.data(), header.cupsBytesPerLine) == 0)
				break;

			//Print in MSB format, one line at a time
			rasterheader(header.cupsWidth, 1);
			unsigned char current=0;
			int bits=0;

			for(int i=0; i < (int)buffer.size(); i++){
				
				//The actual pixel value with gamma correction
				double pixel = pow(buffer[i]/255., 1./2.2) + errors[0][i];
				double actual = pixel>.5?1:0;
				double error = pixel - actual; //This error is then distributed


				//Diffuse forward the error	
				for(int r=0; r < (int)diffusion_coefficients.size(); r++)
					for(int cc=0; cc < (int)diffusion_coefficients[0].size(); cc++){
						int c = cc - diffusion_coefficients[0].size()/2;
						if(c+i >= 0 && c+i < (int)buffer.size() && diffusion_coefficients[r][cc]){
							errors[r][i+c] += error * diffusion_coefficients[r][cc] / diffusion_divisor;
						}
					}

				current |= (pixel<0.5)<<(7-bits);
				bits++;
				if(bits == 8){
					cout << current;
					bits = 0;
					current = 0;
				}
			}
			if(bits)
				cout << current;

			
			//Roll the buffer round.
			std::rotate(errors.begin(), errors.begin()+1, errors.end());
			for(auto& p:errors.back())
				p=0;
			
	
		}

		/* finish this page */
	}
	cout << "\n\n\n";
	cupsRasterClose(ras);
}

And here’s the result

KITTY!!!!!!!!!!!!!

But can we do better? I’m not sure, but look at this:

You can draw on the paper using a finger nail. The faster you move at a given pressure the darker the line. I believe this is due to getting more heating. So, the paper is definitely analogue. Turns out the printer is too, kind of in that you can set the heat output per line (though not per pixel). The command is on page 47 and is the general control command. So what I did is print 255 solid lines, each one with a different heat output. The code (in AWK) is:

BEGIN{
	for(i=0; i < 255; i++){
		printf("%c7%c%c%c", 27, 64, i, 2)
		printf("\x1dv0\0%c\0\x01\0", 40)
		for(j=0; j < 40; j++)
			printf("\xff")
	}
	print "\n\n\n\n"
}

What got me going for ages is that the locale wasn’t C, to characters above 127 were getting mangled. Anyway the results is this:

That’s a yes! It’s a bit speckly, but it can definitely output greyscale. After a bit of messing around, I got the range. It goes from a timing (heat output is essentially controlled by setting the time the heating elements dwell on the paper) range of about 16 (full white) to 112 (full black) and empirically, raising the input to a power of 2 makes it look a little better. Working it into the dithering code is pretty straightforward: find the darkest pixel and set the black level to be able to reproduce that.

#include <cups/raster.h>

#include <iostream>
#include <vector>
#include <array>
#include <utility>
#include <cmath>
#include <algorithm>

using std::clog;
using std::cout;
using std::endl;
using std::vector;
using std::array;


constexpr unsigned char ESC = 0x1b;
constexpr unsigned char GS = 0x1d;

// Write out a std::array of bytes as bytes.  This will form the basis
// of sending data to the printer.
template<size_t N>
std::ostream& operator<<(std::ostream& out, const array<unsigned char, N>& a){
	out.write(reinterpret_cast<const char*>(a.data()), a.size());
	return out;
}

array<unsigned char, 2> binary(uint16_t n){
	return {{static_cast<unsigned char>(n&0xff), static_cast<unsigned char>(n >> 8)}};
}


void printerInitialise(){
	cout << ESC << '\x40';
}

// enter raster mode and set up x and y dimensions
void rasterheader(uint16_t xsize, uint16_t ysize)
{
	// Page 33 of the manual
	// The x size is the number of bytes per row, so the number of pixels
	// is always a multiple of 8
	cout << GS << 'v' << '0' << '\0' << binary((xsize+7)/8) << binary(ysize);
}


void set_heating_time(int time_factor){
	// Page 47 of the manual
	// Everything is default except the heat time
	cout << ESC << 7 << (char)7 << (unsigned char)std::max(3, std::min(255,time_factor)) << '\02';
}

constexpr array<array<int, 5>, 3> diffusion_coefficients = {{
		{{0, 0, 0, 7, 5}},
		{{3, 5, 7, 5, 3}},
		{{1, 3, 5, 3, 1}}
}};
constexpr double diffusion_divisor=42;


double degamma(int p){
	return pow(p/255., 1/2.2);
}

int main(){

	cups_raster_t *ras = cupsRasterOpen(0, CUPS_RASTER_READ);
	cups_page_header2_t header;
	int page = 0;

	while (cupsRasterReadHeader2(ras, &header))
	{
		/* setup this page */
		page ++;
		clog << "PAGE: " << page << " " << header.NumCopies << "\n";
		clog << "BPP: " << header.cupsBitsPerPixel << endl;
		clog << "BitsPerColor: " << header.cupsBitsPerColor << endl;
		clog << "Width: " << header.cupsWidth << endl;
		clog << "Height: " << header.cupsHeight << endl;

		// Input data buffer for one line
		vector<unsigned char> buffer(header.cupsBytesPerLine);
		
		//Error diffusion data
		vector<vector<double>> errors(diffusion_coefficients.size(), vector<double>(buffer.size(), 0.0));

		clog << "Line bytes: " << buffer.size() << endl;
		printerInitialise();

		/* read raster data */
		for (unsigned int y = 0; y < header.cupsHeight; y ++)
		{
			if (cupsRasterReadPixels(ras, buffer.data(), header.cupsBytesPerLine) == 0)
				break;

			
			//Estimate the lowest value pixel in the row
			double low_val=1.0;
			for(int i=0; i < (int)buffer.size(); i++)
				 low_val = std::min(low_val, degamma(buffer[i]) + errors[0][i]);
			//Add some headroom otherwise black areas bleed because it can't go
			//dark enough
			low_val*=0.99;

			//Set the darkness based on the darkest pixel we want

			//Emperical formula for the effect of the timing
			double full_white=16;
			double full_black=16*7;
			set_heating_time(pow(1-low_val,2.0)*(full_black-full_white)+full_white);

			//Print in MSB format, one line at a time
			rasterheader(header.cupsWidth, 1);
			unsigned char current=0;
			int bits=0;

			for(int i=0; i < (int)buffer.size(); i++){
				
				//The actual pixel value with gamma correction
				double pixel = degamma(buffer[i]) + errors[0][i];
				double actual = pixel>(1-low_val)/2 + low_val?1:low_val;
				double error = pixel - actual; //This error is then distributed


				//Diffuse forward the error	
				for(int r=0; r < (int)diffusion_coefficients.size(); r++)
					for(int cc=0; cc < (int)diffusion_coefficients[0].size(); cc++){
						int c = cc - diffusion_coefficients[0].size()/2;
						if(c+i >= 0 && c+i < (int)buffer.size() && diffusion_coefficients[r][cc]){
							errors[r][i+c] += error * diffusion_coefficients[r][cc] / diffusion_divisor;
						}
					}

				current |= (actual!=1)<<(7-bits);
				bits++;
				if(bits == 8){
					cout << current;
					bits = 0;
					current = 0;
				}
			}
			if(bits)
				cout << current;
			
			//Roll the buffer round.
			std::rotate(errors.begin(), errors.begin()+1, errors.end());
			for(auto& p:errors.back())
				p=0;
	
		}

		/* finish this page */
	}
	cout << "\n\n\n\n\n\n";
	cupsRasterClose(ras);
}

And it works!!

Spot the difference! Left: image with enhanced greylevels, right standard image. View image to get the full resolution.

OK, the results aren’t spectacular, but look if you hear a dog talking, you’re impressed that it can talk at all, not disappointed that it can’t talk well.

The enhanced grey level image definitely has some horizontal streaking. I don’t know if that’s due to the printer or the really ad-hoc calibration of grey levels that I did. I should probably limit the rate at which the temperature changes vertically to mitigate that.

Overall I’m really pleased. The finer details are clearer and there are definitely some whiskers which you can make out in the left image which are washed out in the speckle right one. Bare in mind there are optimistically 64 distinct grey levels this printer can produce which means this technique is adding about 6 bits per line of 384 bits.

This also pushes the printer far, far beyond what it was ever supposed to do. The heating time is really a way to reduce print time and/or save on total energy draw, presumably for battery powered chip and pin machines.

I expect there is more fiddling to do, but the next stage is to integrate it into CUPS so I can print the usual way (i.e. using lp of course).

Review: Pi-EzConnect Terminal Block Breakout HAT

I bought one of these recently:Pi EzConnect Adafruit product 2711

I’ve done a moderate bit of raspberry Pi prototyping on and off. The choice has usually been either a piece of stipboard and some cables or an incidental prototyping area on an expansion card of some sort (like the excellent motor hat). This time, I decided to go pro and buy a ready made breakout board.

This one:

2711-06

It’s OK, and there’s a good chance I’ll use it in the final project, but it has a few flaws.

If you look in the photo above, the GPIO holes are doubled up. I can see why they did this but it makes the prototyping area painfully small. In addition, the labels for the terminals and holes is on the inner side. They could have put them on the otherwise outside and had an extra two rows of holes, giving 10 rows instead of 8.

Another flaw is that the painfully small prototyping area is not matrix board, it’s stripboard style. Now normally I’d say this is down to personal taste. I quite like stripboard, personally. However out of the main options of stripboard, tri-pad and matrix, it’s the least compact because of all the track cuts needed. Matrix board is arguably the most faffy and annoying style, however it’s by far the most dense and in a space limited situation like this, density is I think the most important thing.

So essentially, they could have nearly doubled the useful prototyping area of the board, and used a higher density proto typing style.

The other thing I really didn’t like was the terminal style.

20191201_160520.jpg

That leaf spring inside the terminal greatly reduces the usable area, to the point where I couldn’t fit in a piece of tinned speaker wire. I have a 100m reel of the stuff that I use for many miscellaneous wiring odd jobs., and I had to older on a thin piece of single core wire. In fact I have very little multicore cable narrow enough to fit in those holes.

It’s a shame because there are better styles that have a moving gate inside which has a lot more capacity for the same pitch, like this one:

r7176634-01

All the usual suspects have similar types (Phoenix Contact, TE Connectivity, RS Pro, plus a bunch of vendors on Ali Express).

It seems like a selection of small things but over all they add up to a breakout board that’s not as good as it could be. If you have small wires and little soldering needs then it will work perfectly, but it could be a good deal more flexible.

Overall, 3/5

BMVC 2019

I went to BMVC this year and had a great time, and saw lots of interesting papers and talked to a lot of interesting people. BMVC was my first conference in 2003, and it has changed fair bit since then. I remember some of the hushed, awed tones about how was that it was getting really international because there were two speakers from American and one all the way from China. Now it really is a big international conference, that just happens to be located in the UK each year. I think the best bits of the fundamental character haven’t changed.

On the minus side, I made a bunch of notes and then lost them so I’m having to go on memory and have almost certainly forgotten some that stood out. So here’s a somewhat random selection of papers that caught my eye as interesting for various reasons.

But first, here’s a video of Cardiff Science Library vomiting rainbows:

 

A random selection of interesting papers

Keynote

Dissecting Neural Nets
Prof. Antonio Torralba (MIT)

That keynote was very interesting and Prof. Torralba is a fantastic presenter and the results were very intersting. Unfortunately I can’t find the video to link to.

Geometric vision

Whenever there’s a paper not about deep learning there’s always a cluster of people people who’s student days are long past hovering around commenting about how it’s nice to see something that isn’t deep learning. I also like to refer to this type of vision as “geometric” since it involves geometry rather than “traditional” or (even worse) “old-fashioned”.

26. A Simple Direct Solution to the Perspective-Three-Point Problem
Gaku Nakano (NEC Corporation)

The paper is a new solution to the P3P problem. Given the age of the field and number of existing solutions, it’s surprising that there are actually new ones. It’s a surprisingly tricky problem as anyone who’d tried to derive a solution will know and it’s interesting to see there are still new insights to be had.

Adversarial attacks

If you hand a vision system to a computer vision researcher, the first thing they will do is try and break it. These days that’s even publishable!

Non deep image features are still widely used for solving geometric problems especially if efficiency is key. While it’s not surprising, it had never occurred to me that they could be attacked just like neural nets can be attacked.

27. Adversarial Examples for Handcrafted Features
[Supplementary]
Muhammad Latif Anjum (NUST); Zohaib Ali (NUST); Wajahat Hussain (NUST – SEECS)

Much like the attacks on DNNs, the differences aren’t visually apparent. Speaking of adversarial attacks, I found this paper and poster enjoyable and easy to follow, with good results:

210. Robust Synthesis of Adversarial Visual Examples Using a Deep Image Prior
Thomas Gittings (University of Surrey); Steve Schneider (University of Surrey); John Collomosse (University of Surrey)

I didn’t know that was a thing

I like papers that have “towards” in the title. It’s an admission in the title that the results aren’t spectacular or and they aren’t aceing the current benchmarks, but they’re tackling a hard problem in a new way. That’s a good goal for research, not engineering polished solutions, but tackling new problems or bringing new insight to bear.

In this case, they are dealing with point clouds of the sort that might be the result of structure from motion but where the original images aren’t available. Turns out it’s possible to do semantic segmentation of those clouds.

252. Towards Weakly Supervised Semantic Segmentation in 3D Graph-Structured Point Clouds of Wild Scenes
Haiyan Wang (City University of New York); Xuejian Rong (City University of New York); Liang Yang (City University of New York); YingLi Tian (City University of New York)

Realtime semantic segmentation

There’s a lot of interest in realtime techniques  which I like. A lot of it comes from the self driving car industry and all of these are tested on Cityscapes. I’m more interested it from the perspective of running on a phone, but there’s a lot of common ground and so these are well worth a closer look.

253. Fast-SCNN: Fast Semantic Segmentation Network
Rudra Poudel (Tosihiba Research Europe, Ltd.); Stephan Liwicki (Toshiba Research Europe, Ltd.); Roberto Cipolla (University of Cambridge)

259. DABNet: Depth-wise Asymmetric Bottleneck for Real-time Semantic Segmentation
Gen Li (Sungkyunkwan University); Joongkyu Kim (Sungkyunkwan University)

260. Feature Pyramid Encoding Network for Real-time Semantic Segmentation
Mengyu Liu (University of Manchester); Hujun Yin (University of Manchester)

Benchmarks are useful, but I feel that over reliance on them can essentially lead to reverse engineering the datasets. I’ve certainly noticed in my own work that networks that give stellar results on ImageNet don’t do nearly so well when images that aren’t of the sort one posts to the internet (i.e. worse, less well composed, more cluttered, worse lighting and focus etc).

I think all good benchmarks are doomed to eventually become more of a hindrance than a help because of all the focus that they draw. This isn’t to disparage the benchmarks, at all, I think it’s simply part of the cycle of research. I wonder when we’ll reach that point with Cityscapes.

Domain transformation

The key idea here is that (for object detection from a car), a data volume aligned with the ground plane and front of the car is more semantically meaningful than a 2D image. So they transform RESNet features into that cube using a simple technique and do the deep learning there. Sound idea with good results.

285. Orthographic Feature Transform for Monocular 3D Object Detection
Thomas Roddick (University of Cambridge); Alex Kendall (University of Cambridge); Roberto Cipolla (University of Cambridge)

Binary networks

Binarised networks have an appealing minimalism, especially from a hardware and wire-format compression point of view. Unfortunately they’re not differentiable. This paper makes judicious use of carefully inserted weighting factors and derivatives of effectively a blurred binary activation function to introduce differentiability.

19. Accurate and Compact Convolutional Neural Networks with Trained Binarization
Zhe Xu (City University of Hong Kong); Ray Cheung (City University of Hong Kong)

A different approach to deep features

I couldn’t decide how much I like this paper because I kept vacillating about the core idea. Then I realised that in itself makes it a good paper because it’s made me think a lot about the problem. It was very well presented and the core idea is simple and intriguing.

32. Matching Features without Descriptors: Implicitly Matched Interest Points
[Supplementary]
Titus Cieslewski (University of Zurich & ETH Zurich); Michael Bloesch (Deepmind); Davide Scaramuzza (University of Zurich & ETH Zurich)

I like that the features are defined purely by matchability and localisation. I also like that they do not have to do things like have precisely (or at most one) feature per 8×8 (etc) window of the image, and they have a simple structure without auxiliary losses, and an overall simple training procedure.

This is also one of the things I like about BMVC: the results presented in the paper don’t present it as the new leading feature detector, in fact it’s not even near the top of the pack of the ones they compare to. However they’re tackling it in a new and interesting way and I there is a great deal of value in such ideas being shared and discussed even if they’re not (yet?) as good as the competitors.

 

P0533 will ride again

Unfortunately, P0533 (see here for previous posts) didn’t make it into C++20 either (originally targeted at C++ 17). It seems that there were just too many good papers and they couldn’t work through them all in the available time.

There’s lots of good stuff and clearly a strong and growing interest in constexpr’ing everything that can be constexpr’d, so I hold out hope for both it and P1383 in C++2.. uh… 2b? Or not 2b?

Follow their progress in the trackers here:

Light chasing robot part 2 (of 2)

The first version worked, but oscillated a lot in its motion. If you haven’t read it yet, I recommend reading it first otherwise this post won’t make as much sense. And if you have, it might be worth a re-read, since it took me nearly two years to post the followup.

The reason for the oscillation is that it has essentially very high feedback. If it’s very slightly off to one side, then the opposite motor comes on full, because the direction sensor divider goes into a simple comparator. Also, it turns out (I found this about a year later–yes I am a bit lazy about writing blog posts) the response of the LDRs is really slow, measurable over the timescale of a second, so the robot will swing round a significant amount before the resistive divider starts to respond. Either way making the response have a much lower gain will help.

I can reduce the gain by making the motor come on at a reduced speed in proportion to the ratio between the two LDRs.

The circuit is a little more complex than the previous one. It also falls into the category of “should have used a microcontroller” since then the upgrade would just be software and a lot more flexible. Essentially I have used a CMOS 555 in equal duty cycle mode and I’m using the capacitor voltage to get a sawtooth wave. That’s thresholded  by the comparator (opamp) to make a PWM signal. I could have also used the other amplifier in the dual opamp chip to do the same job. That would have been neater in hindsight.

Snapchat-1093110546

Simple PWM circuit

 

The result is really pretty good! See:

 

Er… take 2!

That works well, and is a good validation of the directional light sensors (the original point of this project).

Self feeding flat bits

In knocking together a case for something holding a Raspberry Pi, I needed to cut a 24mm hole for some of these:

f8621579-01

USB bulkhead

My usual go-to stockists didn’t have a 24mm flat bit (or Forstner), so I went to ebay and had a quick dig around. I found a Bosch “Self Cut” spade bit for cheap (maybe not used, old). It looks like this:

Snapchat-926218252

Bosch self-cut bit

Bosch is one of those respectable brands and you won’t go wrong with Bosch tools if you pick the right one for the job.

Speaking of that…

Turns out self feeding bits are wildly unsuitable for the kind of things I do most of the time. They are flat bits but the tip is a screw so it feeds itself into the wood. This should  gives a very consistent depth of cut and chip load. It also means you don’t need to apply any pressure with the drill: it applies an immense amount of drilling pressure using the drill’s torque instead.

They are amazingly, astonishingly aggressive and will happily plough through thick birch ply in seconds (if your drill is up to the task; the level of torque required is vast), and completely split a piece of pine. Note the rather rough cuts with the large amount of tear out:

Holes drilled with self feeding bit

It took much less time to go through the thick birch ply than it took me to drill the holes with a normal bit in a thinner piece of pine.

Great tool, utterly the wrong one for the task.