After 360 commits my coding frenzy has reached a conclusion: ROM v1 is feature complete! The kit will ship not with one but two fast paced games: Snake and Racer. Sound improved and the serial loader is reliable, which is great for hacking and making more games. The fractal program now doubles as a clock, to give a valid excuse for keeping a unit on permanent display in the living room.

The unused ROM space is filled with three classic hires images from my previous projects. By packing 4 pixels in 3 bytes I got three images in where the ROM otherwise would only have space left for two.

This doesn’t mean the to do list is empty, far from that: “make todo” lists 90 ideas I apparently still have in mind. But after 6 months of breadboard prototyping, 3 months of PCB design and 4 months of software hacking, this is a good point to shift focus again. For example, towards demonstrations, tutorials and documentation. Keep you posted.

Lots has happened. Yesterday evening, the software has reached the alfa testing stage. You can find the source in Github. (You can use the visualizer that Martin Sedlák made to run it on your PC while you are waiting for your own Gigatron.)

The assembly videos are now available on YouTube. It is a series, in which I go through all the steps to build a Gigatron. These steps are also explained in detail in the assembly manual that comes with the kit.

At the Hacker Hotel conference, I talked about the Gigatron and people were excited about it! We are excited too, as we are reaching the point where we can actually sell a batch of Gigatron kits. If you want to be informed, subscribe to our mailing list.

While Marcel is busy with the software, meanwhile I am working on over an hour of video footage on how to build the Gigatron. Actually, the 48-page assembly manual that is supplied with the kit already contains all that is needed to understand the electronic components and how to place them, how to solder and how to test the Gigatron during the build process, but we like to make this system as easy as possible to build.

The software to be included with the kit release later this month is nearly done. Just in time as we’re now also waiting for the last parts shipment to arrive, expected in two weeks. When those are good we know if we can meet our target selling price and will announce it to those interested on the mailing list. All other parts are in house already, manuals printed, packaging ready and beta tests successfully completed. Our living rooms look like a warehouse now.

Of course the kit will ship with some demo applications built-in. My focus is for a part still on those, but equally important is that the programming core is stable and tested. For me it is crucial that the memory map is well-defined and the 16-bit interpreter is fully tested and useful. After all, the vCPU opcodes are jump offsets, so it will be impossible to fix any of that later while maintaining compatibility as well. Last week I found I had some unused space in the interpreter code page, so I added some new bit-wise logic instructions and support stack variables. Surprisingly, none of the applications I wrote so far needed those.

To test, I ported my old n-Queens solver. It exercises bitwise logic and recursion, so it is a good test for these new instructions. On the screenshot above you can see it gets the correct answers for the sequence. The solver uses 5 stack variables. That, plus one for the the return address, gives 12 bytes per invocation. With the stack living in the top half of the zero page, this means we can go 10 levels deep. The solver needs less than a minute to compute the list, or at about 850 recursions per second. [Edit: I just figured that it should be easy to go down to using 4 variables or 10 bytes, and with that up to 12 levels deep.]

Although 8-bit assembly programs must be programmed in the EPROM, interpreted programs run from RAM. The built-in applications are of course stored in ROM also, but they are loaded into RAM first, so they use the ROM merely as a disk. Interpreted programs can also be loaded into RAM directly over the input port, and this is how you can program the Gigatron without using an EPROM eraser and programmer. For this I hook up the input port pins, that normally go to the game controller, to a simple Arduino. The Arduino can send data at the same rate as the horizontal sync. With some checksumming overhead, this boils down to exactly 28k8 payload bits per second. Much faster than loading C64 programs from tape back in the day… (3000 baud with speed loaders!)

The loader was the last part of the software that needed debugging with a scope.

I got a bit stuck with the work I was planning for today, so I wrote something else to regain motivation: a fractal! Rendering takes a while, being fully interpreted and lacking multiplication, and even without the “right-shift” operation you badly need for this. All those must be mimicked with slow high-level code using just addition and subtraction.

It should be easy to speed the whole thing up a lot just by adding a right-shift assembly function. Next time… GCL source code:

{-----------------------------------------------------------------------+
| |
| Mandelbrot fractal |
| |
+-----------------------------------------------------------------------}
gcl0x
{
Plot the Mandelbrot set
- 160x120 pixels and 64 colors
- Faithful translation of mandelbrot.c pre-study
- Use 16-bit vCPU math as 7-bit fixed point arithmetic (1.00 -> 128)
- Implement multiplication in interpreter
- Implement shift-right in interpreter as well
- A bit slow (8242.655 seconds)
XXX At the end change all to grey tones and redo
XXX Redo at different sections
XXX Tone for every pixel value
}
{-----------------------------------------------------------------------+
| RAM page 3 |
+-----------------------------------------------------------------------}
$0300:
{ Pretty accurate multiply-shift ((A*B)>>7), but it can be off by one }
[def
push
{Extract sign and absolute values}
0 sign= C=
{0}A- [if>0 A= 1 sign=]
0 B- [if>0 B= sign 1^ sign=]
{Multiply}
7 shift= {Pending shift}
$200
[do
bit=
-$4000 C+ [if<0 C C+ C= else {Shift prematurely in an attempt to avoid overflow} B ShiftRight! B= shift 1- shift=] {Add partial product} A bit- [if>=0
A=
C B+ C=]
bit ShiftRight! if<>0loop]
{Shift}
[do
C ShiftRight! C=
shift 1- shift= if>0loop]
{Apply sign to return value}
sign [if<>0 0 C- else C]
pop ret
] MulShift7=
{ Calculate color for (X0,Y0) }
[def
push
0 X= XX= Y= YY= i=
[do
i 1+ i= 64^ if<>0 {Break after 64 iterations}
{Mandelbrot function: z' := z^2 + c}
X A= Y Y+ B= MulShift7! Y0+ Y= {Y = 2*X*Y + Y0}
XX YY- X0+ X= {X = X^2 - Y^2 + X0}
{Calculate squares}
{X}A= B= MulShift7! XX=
Y A= B= MulShift7! YY=
-$200 XX+ YY+ if<0loop] {Also break when X^2 + Y^2 >= 4}
i
pop ret
] CalcPixel=
{-----------------------------------------------------------------------+
|}\vLR>++ ret{ RAM page 4 |
+-----------------------------------------------------------------------}
$0400:
[def
push
$7ff p= {Start of video (minus 1 to compensate for 1st step)}
-323 X0= 3 DX= 161 Width= {Horizontal parameters}
-180 Y0= 0 DY= 120 Height= {Vertical parameters}
[do
{Length of next segment, either horizontal or vertical}
DX [if<>0 Width 1- Width= else Height 1- Height=] if>0
[do
len=
{Step in the fractal plane}
X0 DX+ X0=
Y0 DY+ Y0=
{Matching step in video frame}
DX [if<0 p 1- p=] DX [if>0 p 1+ p=]
DY [if<0 -$100 p+ p=] DY [if>0 $100 p+ p=]
63 p. {White while busy here}
{First check if we are inside one of the main bulbs for
a quick bailout (Wikipedia)
(x+1)^ + y^2 < 1/16}
Y0 A= B= MulShift7! YY=
X0 128+ A= B= MulShift7! YY+ 8- [if<0 0
else
{q*(q + x - 1/4) < 1/4*y^2, where q = (x - 1/4)^2 + y^2}
X0 32- A= B= MulShift7! YY+ {q}
A= X0+ 32- B= MulShift7! tmp=
tmp+ tmp= tmp+ tmp= {*4} YY- [if<0 0 else {Otherwise run the escape algorithm} CalcPixel! ]] p. {Plot pixel} len 1- if>0loop]
DY tmp= DX DY= 0 tmp- DX= {Turn right}
loop]
pop ret
] CalcSet=
{-----------------------------------------------------------------------+
|}\vLR>++ ret{ RAM page 5 |
+-----------------------------------------------------------------------}
$0500:
{ Stupid shift-right function }
{ XXX Better make a SYS extension for this }
[def
a= 0 b=
$8000 a+ [if>=0 a= $4000 b+ b=]
$c000 a+ [if>=0 a= $2000 b+ b=]
$e000 a+ [if>=0 a= $1000 b+ b=]
$f000 a+ [if>=0 a= $0800 b+ b=]
$f800 a+ [if>=0 a= $0400 b+ b=]
$fc00 a+ [if>=0 a= $0200 b+ b=]
$fe00 a+ [if>=0 a= $0100 b+ b=]
$ff00 a+ [if>=0 a= $0080 b+ b=]
$ff80 a+ [if>=0 a= $0040 b+ b=]
$ffc0 a+ [if>=0 a= $0020 b+ b=]
$ffe0 a+ [if>=0 a= $0010 b+ b=]
$fff0 a+ [if>=0 a= $0008 b+ b=]
$fff8 a+ [if>=0 a= $0004 b+ b=]
$fffc a+ [if>=0 a= $0002 b+ b=]
a 2& [if<>0 b<++ ] b ret ] ShiftRight= {-----------------------------------------------------------------------+ |}\vLR>++ ret{ RAM page 6 |
+-----------------------------------------------------------------------}
$0600:
{ Main }
[do
CalcSet!
60 \soundTimer. {For debugging}
loop]
{-----------------------------------------------------------------------+
| End |
+-----------------------------------------------------------------------}

As I mentioned in the previous post, I did a talk about the Gigatron at the RevSpace hackerspace. One of the people there was Michai, who has a podcast called CBA podcast, about software, electronics and mechanics. He mentioned the talk in his latest podcast (from 6’41).

Last friday, I gave another lecture about the Gigatron at a hackerspace, this time at Revspace. We got some really good feedback and it was very nice to talk to people that have been working on building their own processor or computer.

The lecture was not recorded, but an earlier version of the talk, that was given a month ago at the Hack42 hackerspace, is. It is included here.

While playing Gigatron Snake at 34C3 we got asked: “how many gates are in there?”. This isn’t the first time someone has asked, so today I finally went through the TTL datasheets and counted little blocks from their logic diagrams. TL;DR: the CPU has 930 logic gates.

And because 34C3 is such an inspiring place, after one day of hacking those gates now happily animate this beginning of a Racer game:

Gate counting isn’t as straightforward as it seems though. RAM and ROM are clear: they each have thousands of logic gates in their word line decoders alone, but those aren’t part of the CPU proper. The clock isn’t part of the CPU either, but that’s just a couple of inverters. For decoding the game controller it isn’t as clear cut: the kit version has a 74HC595 shift register which has roughly 80 gates. But only 10 buffer gates are really needed by the CPU and are directly controlled by it. I fact, on the breadboard version, the input chip is just a 10-gate 74LS244 non-inverting buffer. So I count that as one 10 for the CPU. I also didn’t count the “extended output” register as part of the CPU because that is an extension on top of its primary output.

I did include all other gates that are in the IC packages: An unused module is in the count (there is one unused decoder in the control unit), as well as gates that are hooked to fixed inputs (L or H), even though all of those can be optimised away in VHDL. The 4-bit adders each have 4 non-inverting buffers internally that have no logic function, but I still include them. Furthermore, I count D-type flipflops as 5 gates, and SR-types as 4. Still, the total gate count of 930 is a lot lower than I would have guessed.

We are in the process of turning the Gigatron in a kit that can be built by others. For this, I’ve made a full-colour booklet that contains a generic explanation about the electronic components used, a crash course in soldering and the actual instructions to build the Gigatron. It includes many pictures and describes how to perform tests during the building process. In the appendix, you will find the full hardware schematics.

With this manual, even less experienced people should be able to build their own Gigatron.

To make 64 colors, every pixel occupies 6 bits of a byte in RAM. The two high bits can have arbitrary values as they are masked by the pixel burst loop. This opens up all kinds of tricks. For example, they can encode play field information for a game: where there are walls, independent of their color. Here I use these bits in the snake segments: they keep track of direction so the tail can follow the head. Like invisible breadcrumbs. Up/down/right/left just needs 2 bits…