Files

142 lines
4.9 KiB
Markdown

## To build:
``` <!---sh-->
make all
```
## To use:
``` <!---sh-->
./davidguy ip_address:server.screen
```
where `ip_address` in an IP address of an X server.
## Try:
``` <!---sh-->
./davidguy 127.0.0.1:0.0
```
Of course, you may need to supply a more reasonable address. :-)
## Also try:
``` <!---sh-->
cp -f davidguy guydavid
./guydavid 127.0.0.1:0.0
```
Can you determine why this makes a difference?
## Judges' remarks:
We permitted this type of entry to win because:
- it did not require X to be on the host system
- both BSD and SVR4 have sockets
Now that X11 and sockets have become 'standard', we plan to open
up the contest to programs that use them. Future rules will
explain this further.
We are pleased that this entry has helped bring new 'life' into
the contest.
## Authors' remarks:
The program is a fully-functional X client. It talks directly to
the X server through a socket without using Xlib, Xt or any other
wussy toolkit junk. With no arguments, it will try to run
locally using the unix-domain socket (only on BSD :-( ) or you can
give it a command line argument identifying a display host address
(in dotted octet notation, like 127.0.0.1:0.0) where the X server
is running. It handles color and monochrome displays, and
performs whatever byte-swapping and bit-reversal is needed (so
that the program works even when the server and the client have
different architectures). It buffers communication to the X
server for efficiency.
**NOTE**: The original source in davidguy.c used only an IP address
argument (assumed :0.0) and assumed the color of black
and white.
The program doesn't require any user interaction; you can just sit
back and watch it run. All the action takes place in the root
window of the display.
We think the program is interesting for these reasons:
1. It belongs to a class of programs that programmers make the following
complaints about: they require a mountain of code to implement even the
simplest program; and they must be linked with giant libraries of
hard-to-understand functions. In contrast to these complaints (made by wimps,
in our humble opinion), our program fits in under 1536 bytes, doesn't need any
external functions except for a few system calls, and doesn't need to include
any header files.
2. It implements an extremely efficient algorithm for the computation it
performs. One cool feature of the algorithm (at least as far as this contest is
concerned) is that its obscurity is fundamentally necessary for its efficiency.
In other words, even if we took steps to present the algorithm clearly, (we
haven't) the relation of the computation performed to the specification of what
the program was supposed to do would still be hard to understand.
### NOTICE to those who wish for a greater challenge:
**If you want a greater challenge, don't read any further**:
just try to understand the program via the source.
If you get stuck, come back and read below for additional hints and information.
### How this entry works:
The program plays [Conway's Game of
Life](https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life) in the root window's
background pixmap. It starts by setting the background to random
bits, and then plays Life, with one Life cell for each pixel of the
screen. In the game of Life, a cell survives to the next round if
2 or 3 of its 8 neighbors are also alive, and a cell is born if
exactly 3 of its neighbors are alive. Otherwise, a cell dies of
exposure if it has only 0 or 1 neighbors, and dies of overcrowding
if it has 4 or more neighbors.
The algorithm used to compute the next generation is based on the
observation that a cell's state in the next generation is a boolean
function of its current state, and the states of its 8 neighbors
(i.e., a boolean function from 9 bits to 1 bit). This function can
be computed by a boolean circuit. In addition, intermediate values
computed by the circuit can be shared between neighboring cells,
reducing the number of gates per cell required. These ideas
have been used before, to compute the next generation through a
series of [bit blits](https://en.wikipedia.org/wiki/Bit_blit).
Instead of doing this, we map values in the circuit to bits in registers, so
that the next generation can be computed efficiently within registers,
minimizing memory access. As a result, the computation of the next generation
is performed with about 1.6 instructions per life cell, consisting of .125
memory accesses, .17 shifts, and 1.3 logic operations. The net result is that
the time to transfer the bits to the X server, and for the X server to draw them
on the screen, dominates the time to compute the next generation.
<!--
Copyright © 1984-2024 by Landon Curt Noll. All Rights Reserved.
You are free to share and adapt this file under the terms of this license:
Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)
For more information, see:
https://creativecommons.org/licenses/by-sa/4.0/
-->