In Perl, we have some funny (and often contraindicated) features that are just useful enough not to excise them from the language. One of them is that the sort and grep operators (which reorder and filter lists, respectively) return aliases, rather than copies, of the original array elements whenever possible. Because of this, the original array can be modified through the sorted/filtered result.

my @a = qw/w o r d/;
say sort @a; # 'dorw'
 
# Note that perl sort is not in-place.
say @a; # 'word'
 
# Change index 3 of the sort result, which is 'w', to a 'z'
(sub { $_[3] = 'z' })->(sort @a);
 
# The change is reflected in the original
say @a; # 'zord'

Getting this functionality into a non-builtin sub is not straightforward—I haven’t determined any way to return an array directly that retains its linkage to the original. However, it’s pretty easy to do something almost as good (or possibly better): Return an array reference that can do the same thing. The trick to this is that @_, the parameters array for a function, consists of aliases to its parameters, but a reference to it can be taken and returned.

sub parameter_alias_array_ref { \@_ }
 
sub dumb_alternative_sort {
	parameter_alias_array_ref(sort @_);
}
 
my @a = qw/w o r d/;
my $s = dumb_alternative_sort(@a);
say @a; # 'word'
say @$s; # 'dorw'
$s->[3] = 'z';
say @a; # 'zord'
say @$s; # 'dorz'

One place this might be useful is in tandem with the so-called “Schwartzian transform”, a Perl idiom for sorting a list based on some projection of the elements rather than the elements themselves. The premise is that, for some function foo(), we would like the result of

sort { foo($a) cmp foo($b) } @array

but, foo() being somewhat expensive, we would rather compute its value for each element only once rather than the multiple times that would be done this way. So, each row is mapped into a 2-tuple containing the value itself and its projection. The sort is performed ordering by the latter value, then the projection is stripped, leaving only the original value. That looks like the following:

# Note that the order is the reverse of what is described above
map { $_->[0] } sort { $a->[1] cmp $b->[1] } map { [$_, foo($_)] } @array

# Intuitive order using temp variables
# Append projection
my @tmp_appended = map { [$_, foo($_)] } @array;
# Sort by projection
my @tmp_sorted = sort { $a->[1] cmp $b->[1] } @tmp_appended;
# Strip projection
my @result = map { $_->[0] } @tmp_sorted;

Note, however, that this result is a copy, not an alias to original elements.

One workaround would be to have the transform return an array of scalar refs to the original array elements. That would look like this:

my @refs =
	map { $_->[0] }
	sort { $a->[1] cmp $b->[1] }
	map { [\$_, foo($_)] }
	@array;
 
# Assign to whatever the 4th element of the result was
${$refs[3]} = 'idk';
 
# Get the results as a plain array copy
my @result = map { ${$_} } @refs;

Ugly and unnatural. But what if we had some way to transform an array of scalar refs into an array aliased to those scalars?

my $ref_result = scalar_refs_to_array_ref(
	map { $_->[0] }
	sort { $a->[1] cmp $b->[1] }
	map { [\$_, foo($_)] }
	@array);
 
# Assign to whatever the 4th element of the result was
$ref_result->[3] = 'idk';
 
# Get the results as a plain array copy
my @result = @$ref_result;
 
# OR, just use @$ref_result directly!

The implementation follows. The fact that @_ aliases the parameters is heavily abused here. The parameters are aliased a few at a time because the dereferenced scalars must be passed directly and explicitly; the obvious shortcut (using map) removes the magic we’re trying to exploit. An alternative would be to construct an explicit list in the form of a string and use eval to alias all of the elements at once. This isn’t an altogether bad idea, but it’s avoided here.

sub parameter_alias_array_ref { \@_ }
 
sub scalar_refs_to_array_ref {
	my $result = parameter_alias_array_ref();
 
	my $count = 0;
 
	# The seemingly straightforward way to do this would be map { $$_ } @_, but
	# that doesn't preserve the aliasing magic that we need.
 
	while(@_) {
		my @part = splice(@_, 0, 16);
		$count += @part;
		my $pa = parameter_alias_array_ref(
			${$part[0]}, ${$part[1]}, ${$part[2]}, ${$part[3]},
			${$part[4]}, ${$part[5]}, ${$part[6]}, ${$part[7]},
			${$part[8]}, ${$part[9]}, ${$part[10]}, ${$part[11]},
			${$part[12]}, ${$part[13]}, ${$part[14]}, ${$part[15]},
			);
		$result = parameter_alias_array_ref(@$result, @$pa);
	}
 
	# Trim excess introduced when the length is not a multiple of the part
	# length above.
	splice(@$result, $count);
 
	$result;
}

EDIT: You might have noticed that up to 15 undefined array elements (specifically, the remaining undefined elements when there are fewer than 16 in the last part) are dereferenced as scalars when the last part of the input is processed. Yet the code still works perfectly. Why?

It’s a semi-obscure bit involving what’s known in Perl as autovivification: If an undefined scalar variable (say, $somevar) is dereferenced as if it were a scalar ref ($$somevar), an anonymous scalar is automatically created and a reference to it is assigned to the variable (something like { my $tmp = undef; $somevar = \$tmp }), and only then is it dereferenced.

Note that this only works if what is dereferenced is a variable; something like ${(undef)} always breaks. In our case, the variable would be $part[n] for any n greater than or equal to the size of the last part; an array element counts as a scalar variable.

Unless you find that the autovivification of up to 15 scalars is consequential performance-wise (it probably isn’t), the code can be left as is. Otherwise, a workaround would be straightforward.

Reflow soldering typically makes use of solder paste, which is fairly easy to work with but has a short shelf life and can only be reasonably purchased in quantity (read: I can’t justify buying any). I performed a very loose experiment this weekend in using the more traditional coiled solder for reflow.

I’d seen two important articles whose advice I sought to combine. One was for the concept of performing the reflow itself using plain solder and a hot air station. I have no hot air station; what I do have is a reflow skillet,  but that article deals in solder paste.

With these concepts in mind, and with a fresh supply of SMT components purchased in anticipation of building my Bus Pirate, I decided on an extremely simple circuit, an LED and resistor in series, and got to work.

Please note that mistakes were made; don’t try any of this until you’ve read the whole thing and know what not to repeat.

I started by scratching out a PCB. I used a box knife instead of etching because I didn’t want to delay the experiment with an uncertainty and, given my more recent failures, it seemed like etching was more likely to be a waste of time than I could allow. In this case, I scratched the outline of a trace roughly 3mm wide across the length of the board, and sliced across it in two places, one for each component.

I colored all of that side of the board except the six points that needed to remain exposed with a Sharpie, thinking it might work as a solder mask. (It didn’t. Read on.)

I used my soldering iron to apply solder to each of the six points I left uncovered; this would be the solder to be reflowed.

Sharpied and soldered. Only one of these was a good idea.

After it cooled enough to handle, I placed the components atop the now-solid solder on their pads. Both components were 0603^[1], roughly the size of a sesame seed, and nudging them around without flipping them over is difficult, especially with my unsteady hands. One also must take care that the part stays in place instead of sticking to the tweezers. This was a trial of patience.

At this point it was necessary to apply flux; the original flux in the core of the solder itself would have already dissipated when soldered the first time. The flux itself exposed the main two errors that I made in this process:

• Components should probably be placed after the flux. I had expected the flux to be liquid like in the videos, but at somewhat below room temperature (the temperature of the workspace), it is most sincerely a paste, and it is easily stiff enough to knock the already tenuously-placed components out of place. Bigger components are probably less of a problem since they are already easier to place.
• Sharpie does not work as a solder mask, at least for reflow; flux readily dissolves the ink. Actually, the ink did seem fairly effective when I was applying the solder with the iron, but it started to lift with the paste, and once the flux was warm enough to flow the whole board was a tacky, blue mess. It naturally follows that the solder wicked all the way up the traces, as evident from the picture, but in this case that wasn’t a problem.

Once everything was relatively ready, I put the board in the pan^[2] and turned it up to 375°F, which is reportedly slightly above the melting point for my solder. The board was placed off-center in the ring-shaped area directly over the heating element. The flux quickly turned liquid. After only a couple of minutes, the solder liquefied as well, and once it all seemed thoroughly reflowed, I turned off the heat, leaving the board in the pan as it cooled^[3].

Nice board, if you like your flux tinted.

I brushed off the excess flux and scratched through the remaining film to the contact spots I made for power to the board. I connected up the gator clips and powered up. Nothing in either direction. Then I realized that I’d left the plane on the far side of the board untouched, causing the clips to be shorted. Whoops.  Touching the clips to the points on the top side of the board allowed it to function properly.

It can be done!

Stuff I’ll keep in mind for next time:

• If more than a little solder ends up on a pad, it should probably be wicked off. A little is necessary; a lot can be problematic for placement.
  • I’ve also read that, for small enough parts, unevenness between the pads can cause the component to stand on end instead of soldering properly, called “tombstoning”. Let’s try to avoid the parts that are especially prone (the hard-to-imagine-ly small 0402 and 0201 parts).
• Flux before placing. It may help keep the components in place.
• Experiment with mask materials before attempting to use one again.

Thanks to Club Cyberia for being a place other than my garage where I can make this stuff happen.

1. ^[1]0.06 by 0.03 inches↩
2. ^[2]To be perfectly honest, the board was in and out of the pan at least twice as I noticed things I’d forgotten to do, but was only allowed to fully heat when ready.↩
3. ^[3]Cooling or quenching it suddenly is likely to damage the components or reduce the structural integrity of the connections.↩

ICSP in general

From the perspective of actual information to be transferred, an ICSP (in-circuit serial programming) programmer for a PIC microcontroller is not terribly complicated. As with a CPU, there are operands and instructions that must be issued in a particular order.

General pinout of PIC ICSP connector

As with a serial shift register or SPI device, there are two lines—one data line (pin 4, ICSPDAT) to carry the information plus one clock line (pin 5, ICSPCLK) to indicate when that information is valid. In ICSP, the data line is bidirectional so that the programmer can read memory from the target PIC; the programmer should be prepared for that.

There are also the Vpp (pin 1) and Vdd target (pin 2) lines. Vpp must be able to assert either 0V or the programming voltage (13V for most targets). Similarly, Vdd must be able to assert either 0V or the target Vdd (often 5V, but may vary). The switching times must be kept to a minimum; for example, a 16F88 target specifies a Vpp rise time of 1µs.

Vss (pin 3) exists but is not usually switched.

Finally, there is AUX (pin 6), a plain output pin from the programmer to indicate LVP (low-voltage programming) instead of supplying the high voltage on Vpp. Programmers not supporting LVP tie this to ground via a resistor.

The application circuit itself, when presenting an ICSP header, must isolate the pins of the header from excessive capacitance or other interference so that the voltages and rise/fall times are as specified. If possible, the pins used for ICSP should be dedicated. If one or more of these pins is needed for I/O in the application, the circuit should be designed to give the ICSP higher priority (such as by connecting the ICSP header directly but connecting the application circuit via a resistor). If Vpp is connected to an RC slow start circuit (as /MCLR), it can also be isolated using a resistor or a Schottky diode.

Production programmers

The documentations for ICSP differentiates a prototyping programmer from a production programmer as follows:

A production programmer is capable of performing the verification step (i.e., reading back the contents of the target device’s memory as written) with target Vdd set to the minimum and maximum specified voltages of the application circuit.

A prototyping programmer…doesn’t.

This verification is intended to ensure that each memory location in question was either written or erased with a sufficient combination of voltage, current, and time to be reliably recovered.

As implemented by Ardpicprog

Ardpicprog is a project that specifies an ICSP-compatible prototyping programmer using an Arduino for the host computer connection and most of the programming logic.

On the one hand, if the application circuit is simple enough, or if you use the ZIF socket configuration instead of ICSP itself, the programmer circuitry could not get much less complicated, constructed only of three resistors and a single transistor, plus whatever 13V supply you have handy for Vpp^[1]. If you’re in a hurry, this could be your pick.

Switching circuitry for Ardpicprog (not shown: 13V linear regulator for Vpp_SRC, indicator LEDs, ZIF socket for non-ICSP use)

However, there are some issues that may prevent it from being generally applicable, at least for ICSP:

• ICSP_Vdd is driven directly from an Arduino I/O line. For non-in-circuit programming and for the very slightest of application circuits this is fine. But the absolute maximum DC current rating for that pin (ATmega168) is 40mA, so the  Vdd pin on the application circuit might need to be extremely well isolated in order not to harm anything.
• ICSP_Vpp has a somewhat high impedance (R2) when outputting high. This might not be a problem in most situations, since the target’s Vpp pin is fairly high-impedance itself, but when programming I noticed that this line was closer to 12V when turned on.
• The programmer’s logic high voltage is the only target Vdd supported.
  • If the programmer is 5V but the application circuit is designed for 3.3V (some PICs can do this), incomplete isolation on the application side could result in damage.
  • If the programmer is 3.3V (some Arduinos and copycats can do this) but the application circuit is designed for 5V, programming would probably fail unless Vdd to the PIC device is well isolated and the PIC is of a model that supports 3.3V (such as 16LF88—note the L).
  • The programmer is not properly buffered for a self-powered target. For one example, a 3.3V programmer with a self-powered 5V target would probably sustain damage from the current sunk by the ICSP_Vdd pin.
  • Naturally, the programmer is not for production.
• LVP is permanently disabled. (This isn’t necessarily a bad thing; I’ve read that actually getting LVP to work is difficult.)

As implemented by PICkit 2

Microchip’s own PICkit 2 is a mostly production-capable PIC ICSP programmer. The designs for this 18F2550-based programmer aren’t free (in the sense of liberty), but the firmware and schematics have both been published and clones of the device are plentiful.

Switching circuitry for PICkit 2

While not nearly as simple as Ardpicprog, the switching for the PICkit 2 is not utterly complex, either.

• The Vpp switch uses a sort of push-pull, allowing both high and low voltages to come through with fairly low impedance. This is controlled by two lines, and the firmware must prevent both from being on at the same time to avoid shoot-through.
• The Vdd switch is similar, but uses MOSFETs, presumably to reduce voltage drop, and a Schottky diode to prevent reverse current from a self-powered target.
  • A self-powered target is supported by turning off both push and pull in firmware.
  • Whether self-powered or not, the actual target Vdd appears on ICSP_Vdd.
• The data, clock, and AUX lines are clamped, using PNP transistors, to be no higher than ICSP_Vdd. (The 18F2550 correctly reads 3.3V-level highs and lows even at 5V).

Perhaps the most interesting characteristic of the PICkit 2 is the onboard boost regulator.

PICkit 2 Vpp boost regulator

Using a PWM output^[2] (PK2_Vpp_PUMP) with an inductor, the programmer is able to produce and regulate the high Vpp voltage (Vpp_SRC) without a second supply. A resistive divider feeds back to an ADC input (PK2_Vpp_FEEDBACK), allowing the firmware to make adjustments to the PWM duty cycle.

This is by no means an original circuit, nor is it specific to PIC (I’ve gotten one working on an Arduino), but it is a useful inclusion.

What makes this a more-or-less production-capable programmer is its built-in adjustment for Vdd.

PICkit 2 Vdd buffer

A PWM output^[3] (PK2_Vdd_TGT_ADJ) is used as a rough DAC by passing it through an RC lowpass (R22 and C2), resulting in a steady analog level. The op amp is configured to double the input, so the DAC must output half of the intended voltage. The PFET in the feedback loop increases the current capacity of the amp output.

The maximum target Vdd is less than the programmer’s Vdd due to drops that exist within the programmer. One workaround is to make the target self-powered; all of the lines are set up such that even a 6V target shouldn’t harm a programmer running at 3.3V^[4]. By self-powering like this (with any programmer), it becomes more difficult to enter program mode with some targets: The timing requirements vary from device to device, but several of them require Vpp to rise a very short time^[5] after Vdd initially rises. When the target is self-powered, ICSP_Vdd cannot be driven low by the programmer. There might be some secret to making this work that I just haven’t read yet.

It would also work to patch in a higher voltage (maybe 6-7V) to the source of the PFET instead of Vdd. For an actual PICkit 2, this would involve some messy hacking that would more likely ruin the device than improve it.

My own ideas

I’ve been toying with an idea to make a production-capable (or nearly-production-capable) programmer based on the principles of these designs plus some of my own ideas.

• Arduino implementation.
  • This is partly because I think it’s funny, but mostly because it caters to impatience: If you suddenly need a PIC programmer, nobody sells a PICkit retail but at least RadioShack and Fry’s have some variation of Arduino for sale.
• Full support for any combination of 3.3V or 5V programmer and target.
  • Some Arduinos and other development boards are pushing for 3.3V logic. I say, sure, why not?
• Push-pull switching for Vpp and Vdd.
  • I think the Vpp switch on PICkit 2 would be a good model for both Vpp and Vdd. I don’t dislike MOSFETs, but I have a zillion PNP and NPN devices and only a very small number of FETs.
  • Using the PK2 Vpp switch for Vdd would allow use of e.g. 5V target with 3.3V programmer. Q2 and Q3 serve as a sort of level shifter.
• Full level shifting for I/O lines.
  • A level shifter I’ve been playing with in a simulator^[6] starts with a buffered 1V supply (2 silicon diodes, 1 Schottky diode, 1 NPN, and input voltage of 2V or higher) through 470 ohms to the base of an NPN. The input is the emitter of that transistor, and the output is the collector, which is pulled up to the desired output voltage.
  • A bidirectional variation of this^[7] is to use 2 NPNs with the bases tied together (to the same 1V through 470 ohms) and cross couple each collector to the other emitter, finally pulling up each side to the desired voltage.
    • Unlike the common 1-NFET bidirectional shifter, this design does not require the advance knowledge of which side will be higher, so the target is allowed to be either higher or lower than the programmer.
  • Since these shifters pull up rather than down, firmware or other circuitry expecting pull-down behavior would need to be adapted.

1. ^[1]Several are listed, the simplest being a 13V linear regulator based on an adjusted 7812.↩
2. ^[2]Configured to 150kHz.↩
3. ^[3]Again, configured to 150kHz.↩
4. ^[4]But don’t quote me on that.↩
5. ^[5]Less than a small number of ms↩
6. ^[6]Not an original creation.↩
7. ^[7]Also not an original creation.↩