Square: 0x07 cycles
Checkerboard: 0x4C cycles
4 Squares: 0x09 cycles
My key insight was twofold:
1. If you start every thread at 0x00 and make the first N commands "THR @00", you get 2^(n+1)-1 threads running by the time your first thread exits the chain of THR commands.
2. You can use a single array MOV instruction to set a range of program counters, effectively causing every thread to jump to a set position.
So in the first 5 cycles I start 31 threads, then my main thread drops into a loop where all it does is set every thread's position (including its own) to a start point every 2 cycles. Every single thread is now a 2-cycle loop with no JMP cycle spent to return to the start each time. Then I simply make 16 loops of:
PIX counter color
ADD counter 1 counter
Note that, due to space constraints, I use the unused 4th byte of each PIX instruction as the counter for that loop.
Doing this got me to 0x24 cycles. I shaved a further 4 cycles by using the leftover space to add 4 more loops. These ones run vertically and operate on the last two columns (two threads doing one column together). By the time the row-moving threads have reached the 2nd last column, the column-moving threads have finished the last two columns, and so the program is finished 4 cycles earlier.
I think I could get another 2-cycle saving if I fine-tuned the number of threads started, compressed everything down again and got another two column-moving threads.
The bigger improvement is converting the whole thing to 1-cycle loops instead of 2-cycle, and having half the loops incrementing the counters and the other half doing PIX instructions. This might just be possible in the available space if I've run the numbers correctly, and would result in a total time of 0x14.
EDIT: Did the latter: http://imgur.com/zD1Gw5H
It came out to 0x16, as I failed to account for the extra time to spin up more threads.
http://i.imgur.com/sD5IAXP.png
The thread sync is pretty inefficient. It will need to incorporate some "tricky" techniques like what you've described to go faster.
As for the language itself, a MOD operation to go with the DIV operation is useful. bitwise AND / OR / NOT / XOR could be interesting.
Something that would make threads a lot more versatile would be some way to have them behave differently from one another even when executing the same code.
Perhaps a new instruction like JTI X Y: Jump if Thread Id (defined as 0xFF - location of program counter) is equal to Y
Or much more versatile but harder to account for in opcodes: a new prefix #XX which behaves like @(XX + thread id). So you could, for example, write 0x42 to a per-thread slot in an array starting at 0xA0 by doing "MOV 042 #A0"
For making the puzzles interesting: The problem is most shapes and images can simply be described in memory and subsequently brute-force PIXed in sequence. After the simplest puzzles, the trick is having images that do not fit easily in memory and require procedural description. Checkerboard is a good example here. What about something like fizzbuzz but described in colors (like this: http://imgur.com/RMEaKAg)? Or tricky questions where a seemingly random image actually has a simple pattern (like this: http://imgur.com/TbTPZnh (spoilers here: http://imgur.com/73De5QF ))? Unfortunately you're very limited as there's no actual input; the program's exact output is always entirely known.
Several people here have mentioned TIS-100: It's not the exact same thing as what you're doing, but it could be an interesting thing to compare with if you haven't played it yet.
EDIT: Just for fun, my best time for that second puzzle I posted is 0x103.
The copy-pasting to clipboard doesn't work in the webgl for browser security reasons, but you can download a Windows build that has working copy to clipboard. I'll make an OSX build later also. See the link here: http://bit.ly/1V1PiHt
How to make threads operate differently on same code is interesting. I don't really want to do another prefix, because I want to keep the realism of 255 opcodes and I've already used 128. The permutations of new prefix would not fit anymore. I also like the JTI suggestion, but perhaps it doesn't go as far as I would like.
Currently I'm thinking that maybe in addition to program counter, the thread would have another slot where it can define "memory offset". You could state the offset when you start the thread, but it is also modifiable later on, as its in memory. You would call for example "THR @023 010 000", which would essentially mean that for ALL operations for that thread, ALL memory addresses get a offset of +010. I think this is very close to what you suggested with threadID. I think the offset would live next to program counter (FE & FF). Do you think that would work?
I've played TIS-100 and named this game similarly BOX-256 as a homage, since TIS-100 pretty much inspired the whole thing. It's a wonderful game.
Great ideas on the new levels. I will shortly update with more levels and I'll be sure to add your suggestions in to the mix :)
The "all operations offset" sounds tricky but interesting. It'd be problematic in a typical usage scenario, but probably work well with my horrible "every line of code runs in a 1-cycle loop on its own thread" style :P
I'll take a step back and think about the threading some more.
Checkerboard and 4 Squares: http://imgur.com/a/dJRhM