Skip to content

ricardoquesada/64k-ought-to-be-enough

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

64K RAM Ought to be Enough

(Updated 2018-10-08)

64K Ought to be Enough

A demo for the IBM PCjr by Pungas de Villa Martelli.

It was presented at Flashparty 2018 and won the Demo category.

Requirements

An IBM PCjr with at least 64k RAM.

Download:

Compile

Prerequisites:

# compile the resources
make res

# test the disk
make test_boot

# from dosbox-x console run:
c:> boot -l a

Technical description

The demo is divided in the boot loader and demo 3 parts

Boot loader

Boot loader

The demo is intended to work with a 64K RAM (or more) PCjr. Booting from its own boot loader is needed to save precious memory. DOS alone takes ~20K of RAM. That is 30% of the total memory. You don't want to waste that memory.

The boot loader is pretty simple:

  • It has a list of sectors to load for each part. Each part is loaded starting at 0050:0100 (which is the same as 0060:0000) and after loading it, jumps to that address.

  • Installs ricarDOS, a mini "DOS" that hooks the 0x21 interrupt handler with:

    • ah == 0x4c:
      • in ricarDOS it terminates current scene and loads the next one.
      • in real DOS it terminates the current DOS program.
    • ah == 0x09:
      • prints a $-terminated string

    ricarDOS was created to have a "rapid testing framework". It allows to run each part on an emulator using a real DOS, but when run from the boot loader, the actual DOS calls are redirected to ricarDOS.

The memory is organized like this:

Memory map:
 0      - 0x01ff: 512 bytes used for the vector table for the first 0x80 interrupts
 0x0200 - 0x03ff: 512 bytes used to store ricarDOS
 0x0400 - 0x04ff: 256 bytes. BIOS variables
 0x0500 - 0x05ff: 256 bytes. stack. used globally for all the demo parts
 0x0600 - 0xffff: 64000 bytes free to be used for the demo (including video)

Part I

Graphic Big fonts

Big Fonts

The big fonts are based on Andrew Glassner's Notebook. Basically the rectangle to draw each font is divided in 55 segments. Think of the typical 8-segment display, but instead of 8, it has 55 segments. With 55 segments you can render pretty nice fonts. Specially if they are designed by Andrew Glassner.

55 segments

So we have 110 (55 + 55) primitives:

  • 55 primitives to turn on each segment
  • and 55 primitives to turn them off

Each letter consists of a 64-bit bitset containing the 55 segments.

The bitset for letter A is:

;ASCII: 0x41
table_a:
    dw 0b1110010011101110,0b1111111101000101,0b1000111111111111,0b0000000000000111

And this one is for letter B:

;ASCII: 0x42
table_b:
    dw 0b1010010011101111,0b0111111011111100,0b1111110001111111,0b0000000001111011

Notice that A and B have many segments in common. If letter B should be drawn right after letter A, then only the "diff" between A and B needs to be drawn. A xor B

And the diff between A and B is just a simple xor between A's bitset and B's bitset.

A: 0b1110010011101110,0b1111111101000101,0b1000111111111111,0b0000000000000111
   xor
B: 0b1010010011101111,0b0111111011111100,0b1111110001111111,0b0000000001111011
   -----------------------------------------------------------------------------
   0b0100000000000000,0b1000000110111001,0b0111001110000000,0b0000000001111000

0 are skipped (it means that the segments are the same), while 1 are processed. The segment is turned on or off accordingly. For A xor B only 18 segments needs to be updated (instead of 55!).

And that's it.

Misc:

  • Aresluna editor was used to get the definition of each letter.
  • We used a custom script to automatically generate all the 110 primitives and bitsets.

Graphic Loading

The Big Font uses the 320 x 200 @ 4 colors video mode. 16K RAM needed is needed for it. So, from the 64000 bytes reserved for the scene, 16k will be used for the video mode. That leaves 47616 bytes free for the code. And the code to render the big fonts takes ~47000 bytes, with only a few hundred bytes free.

In order to add an additional 16K graphics, the new graphic was appended after the Big Font code (with some padding).

part1.com format
+--------------+  0x00600
|              |
| Big Font     |
| code + data  |
|              |
|      |       |
|      V       |
|              |
|  end of      |
|  Big Font    |
|~~~~~~~~~~~~~~|
| padding...   | ~0x0bf00
|~~~~~~~~~~~~~~|
|              |  0x0c000
| graphics     |
| data         |
| (16K RAM)    |
|              |  0x0ffff
+--------------+

So when the part1.com file is loaded, the graphic will be loaded right where the video card expects it. And it will be displayed automatically.

Part II

Graphic

From a technical point of view, nothing interesting really happens in Part II. It is just a simple horizontal scroll that consumes almost all the CPU cycles.

[Note: Additional effects were planned for this part, but we didn't have the time]

It uses a 320x200 @ 16 colors video mode. In order to "enable" this video mode in a 64k-only PCjr you have to do, but see "Embarrassing bug" section for more info.

sub     ax,ax
mov     ds,ax                       ;ds = 0
mov     word [0x0415],128           ;make BIOS set_video_modo believe that we
                                    ; have at least 128K RAM, otherwise it won't let
                                    ; us set video mode 9

Part III

Vector fonts Moon Still there

[Note: Originally, we wanted to add some 3d effects in Part III. But after doing some performance tests with basic 2D polygons, we decided it was not worth it. The PCjr was too slow for what we wanted to do. We reused part of that code for the Vector Fonts.]

Vector fonts

By vector font we mean a letter that is defined by one or more polygons. I'm using a primitive to draw a line called Line08_Draw (code based on Richard Wilton book). On top of that we implemented a draw_poly function. And each letter is defined by a list of polygons.

The benefits of vector fonts, it is to scale them up/down & rotate them for free, without losing precision.

The fonts are defined using polar coordinates (radius + angle); not the more common cartesian coordinates (x + y). This is legacy code from my 2d-polygon experiment. Using polar coordinates + lookup table is a nice and fast way to rotate polygons. But in retrospective, kind of overkill for Part III.

Render buffer

Mode 160x100 @ 16 colors (Trixter's variation) is being used for the vector font part. And the bottom 40 rows are reserved for the fonts. That's a total of 3200 bytes (160*40/2).

We are using an additional 3200-bytes buffer (a render buffer) where the text is pre-rendered. Then the render buffer is copied row by row, one at the time, to the video buffer. The video buffer is scrolled up or down (depending on the effect). And that is basically it.

Easter Egg

Easter Egg

The final part is pure PCjr BIOS code.

  • Populate the keyboard buffer with "PVM RULEZ!"
  • Put sprite data in the correct place
  • Initialize some video variables
  • ...and jump to the correct place in BIOS.

The sprite data needs to be placed at 0060:0000, the same space used for the demo, so extra careful was needed to not overwrite our own init-easter-egg routine.

Misc

All the "official" PCjr video modes were used in this demo, without any repetition.

  • Boot loader: 40x25, 80x25 (when PCjr is not detected)
  • Part I: 160x200 @ 16, 320x200 @ 4
  • Part II: 320x200 @ 16
  • Part III: 160x100 @ 16, 640x200 @ 4, 640x200 @ 2

Embarrassing bug

The demo is all about the 64k-RAM PCjr. The embarrassing part is that some graphics don't look Ok with the 64k-RAM PCjr. The high-density video modes (80x25 text, 320x200 @ 16 colors and 640x200 @ 4 colors) are disabled in the 64K-RAM PCjr for a good reason: the PCjr can't display them, at least, in theory.

Two months before the deadline, we did a test on a 64k-RAM PCjr, and the 320x200 @ 16 colors video mode looked correct... only because we tested with a single color graphic. We painted the screen with one color, and it looked fine.

But one day before the deadline, while we were testing a release-candidate on the 64k-RAM PCjr, we noticed that some graphics weren't looking Ok. We panicked. We couldn't believe what were seeing. We tried to understand what was happening, but we failed and couldn't fix the bug.

Looks OKish from far away but looks jagged from a closer look

So we shipped the demo with this embarrassing bug. A few days after the party, and after discussing it in the PCjr forum, we realized that the 64k-RAM PCjr has the following limitations:

The system board 64K RAM is mapped at the bottom of the 1Mb address space. The system board 64Kb RAM is mapped to the next 64K bytes of address space if the the 64Kb Memory and Display Expansion option is not installed.

When inserted, the memory expansion option uses the ODD memory space, while the system memory space is decoded as the EVEN memory. Thus, when used as video memory, the memory expansion option has the video attributes while the on-board system memory has the video characters. This arrangement provides a higher bandwidth of video characters.

In addition to the eight memory modules, the expansion card has logic to do EVEN/ODD address decoding, video data multiplexing and Card Present wrap.

And that explains why the 640x200@4 graphic looked as a monochrome bitmap and the 320x200@16 looked jagged. (See "PCjr Technical Reference" Section 2-55)

Monochrome when in 64K RAM only

We tried to disable the "even/odd" decoder by setting "Mode Control 1"'s High-Bandwidth register to 0. But that brings other issues. We still don't know whether or not it is feasible to display a high-density graphic in the 64K-RAM PCjr. Although everything seems to indicate that it is not feasible.

The not-so-bad news is that the 128K-RAM PCjr (the most popular model), is as slow as the 64k-RAM version but runs the demo correctly. No glitches in the graphics.In fact, if we had targeted the 128k-RAM version, probably we could have fit everything in a single executable + our boot loader.


Packages

No packages published

Languages

  • Assembly 89.8%
  • Python 9.5%
  • Makefile 0.7%