Where do all the bytes come from? | Hacker News Reader

Where do all the bytes come from?

medium.com

90 pointsEddieRingle27 comments

I don't know, it's a bit like taking a screenshot of a text file, and wondering why the screenshot is 64k and the text file is only 500 bytes (or whatever).

That's a much better analogy.

I was expecting more of a realistic explanation about how the NES puts up the frame without a frame buffer. This falls into a mess about graphics compression, which is kind of irrelevant.

You would like this:

"How 'oldschool' graphics worked Part 1 - Commodore and Nintendo" https://www.youtube.com/watch?v=Tfh0ytz8S0k

I prefer Montfort and Bogost's Racing the Beam from MIT Press:

https://mitpress.mit.edu/books/racing-beam

Indeed. I also didn't much like the comparison to the .kkrieger demo saying it was "only" 24KB more. Sure the raw exe is but it still needs GB of memory and a crazy fast CPU to do all the content generation at runtime.

He also used inconsistent mb/MB, etc. which always gets on my nerves in a technical post but now I am just being an asshole at it is 2am here ;)

Minor nitpick, but it screws up all the calculations:

The original NES console was only designed to output images that were 256 wide by 240 high; meaning that the final image that needed to be displayed to the screen was 180kb in size.

The NES definitely didn't have 24-bit colour, so the final image data was at most 60kb, assuming 256 colours, or 30kb assuming 16 colours and a palette.

I don't know for sure what colour settings the NES had, I doubt it had a freely selectable 256 colours for each pixel. Probably a limited palette, maybe per-sprite, maybe for the whole screen.

> I don't know for sure what colour settings the NES had, I doubt it had a freely selectable 256 colours for each pixel. Probably a limited palette, maybe per-sprite, maybe for the whole screen.

http://www.thealmightyguru.com/Games/Hacking/Wiki/index.php?...

"a total of 64 pre-set colors (56 of which are unique), and you can only show 25 on the screen at one time (under normal circumstances)."

So ya, a little more limited than 24-bit colour.

Wikipedia to the rescue: https://en.wikipedia.org/wiki/List_of_video_game_console_pal...

- 4 palettes for sprites at a time, each containing transparent + 3 colors. So 12 sprite colors.

- 4 more palettes for background (applied to 16x16 areas), each sharing their first color, for another 13 colors in the background.

- Total up to 25 colors on screen simultaneously, selected from a 64 color palette that really had 54 useful colors.

Reminds me of a sarcastic response from my classmate back in my undergrad Signals class... the professor showed us a sample black-and-white photo and asked (rather rhetorically) how we would represent it using as few bits as possible.

Classmate said he would build a machine that only displays that image, and give it an On/Off switch - a whole image with only 1 Bit!

And this is the crucial insight. An image, or a game, does not contain all its data in those bytes you have stored in its file. A lot of information is spread around in the environment. It's one of the main ways demo-makers can cram so much content into so little space - they pull as much stuff from the environment as it's possible (and allowed by contest rules).

To give an example: one of the most basic tricks, if you're allowed to use OpenGL and WinAPI, is to use WGL to render your system fonts into 3D shapes[0]. This allows you to get a lot of 3D models from just the few bytes it takes to call wglUseFontOutlines()[1]. You may think that there isn't anything interesting in a basic font, but - even forgetting about Unicode - every Windows has Wingdings...

(Or look into the good old .EMF[2] format, which was basically direct encoding of Windows GDI calls.)

For much more in-depth discussion on where exactly information is stored, be sure to read GEB[3], if you haven't already.

[0] - https://www.youtube.com/watch?v=iri5CgLglkk - 1/10th size of Mario!

[1] - https://msdn.microsoft.com/en-us/library/windows/desktop/dd3...

[2] - https://en.wikipedia.org/wiki/Windows_Metafile

[3] - https://en.wikipedia.org/wiki/G%C3%B6del,_Escher,_Bach

> And this is the crucial insight. An image, or a game, does not contain all its data in those bytes you have stored in its file. A lot of information is spread around in the environment.

This is also an argument I've heard about growing people new bodies to transfer into (like Old Man's War) or digitizing humans (like the Singularity); apparently, a sample of DNA isn't enough to create a whole new person. You need, at least, a womb.

This also ties back to the Reflections on Trusting Trust[0]. Living things are not made in a single factory, but are akin to the compilers compiling compilers. It's interesting to ask just how many things about us are not expressed in the DNA at all, but instead propagated implicitly by the replication mechanisms, Ken-Thompson-style.

[0] - http://fermatslibrary.com/s/reflections-on-trusting-trust

This is text-book Kolmogorov complexity. https://en.wikipedia.org/wiki/Kolmogorov_complexity

Why have not zero bits then? http://www.dangermouse.net/esoteric/lenpeg.html

> Why have not zero bits then?

LenPEG 2 contains a bug that will cause it to corrupt images.

Let's say I have a custom compression algorithm (as allowed by LenPEG 2), let's call it "SurprisedBabyPEG"; it's steps are:

1. Is the image the "surprised baby gif"? If yes, write an empty file.

2. Otherwise, encode the input as PNG, and write that to the file as output.

LenPEG 2 allows us to choose an encoding algorithm:

> 2. Prompt the user to ask what other algorithm to use (GIF, JPEG, PNG, etc).

Take the surprised baby gif, and feed it as input to LenPEG 2. It's not Lenna, so step 1 fails. We proceed with step 2, and it prompts us for an encoding algorithm; we choose "SurprisedBabyPEG".

> 3. Use the user-suggested algorithm to compress the image and write it.

SurprisedBabyPEG outputs the empty file; this is the same output as if we run Lenna through LenPEG 2; on decompression of our surprised baby, Lenna is incorrectly output.

(This bug is correctable, and also affects LenPEG 3.)

Also,

  (grep "`./smr`" < input || cat lenna) | xv -

The linked smr.c does not link.

This reminds me how you can communicate any finite sequence of octets across a network connection with only a two bits.

Doesn't this bug affect LenPEG 1 then?

An compression algo that returns "surprised baby gif" if it has a file with first bet set to 1... would match LenPEG 1.

- The tiles have many repeated images while in the original they could flip them and change the color palette. Note that the bushes and the cloud share sprites.

- .kkrieger wasn't 64 kb, it's 96 kb. Still impressive though.

- The image in the tweet is 51.4 kb.

I was expecting some explanation about how the sprites are so tiny because they use a palette of 4 colors (and 4 pixels fit in one single byte), then the colors themselves are stored separately.

I feel like this is pretty important since the article is talking about the comparison in the first place.

Here's a couple blog posts on the subject, which I think explain it better.

http://www.dustmop.io/blog/2015/04/28/nes-graphics-part-1/

http://www.dustmop.io/blog/2015/06/08/nes-graphics-part-2/

I don't see why it's an "unfair" comparison. You just need to think of it in the right context.

At the time Super Mario was popular, storing a still raster image took a lot of hardware. Generating an interactive picture on a TV was easy enough on consumer hardware, but simply recording a picture was not! Today we take the concept for granted, eg digital cameras.

I was mildly surprised when John Carmack retweeted the original tweet. https://twitter.com/smashingmag/status/675624576630571009

The state of the image (jpeg artifacts), was a dead giveaway that the comparison is worthless.

> a dead giveaway that the comparison is worthless.

Worthless to anyone who understand both how the NES and JPEGs work, but pretty worthwhile to someone who's just learning about it.

Yea, and the reference links were interesting too. The LZSS youtube video was an enjoyable watch.

To be fair, the original tweet references "website weight" which is dependent on generic image formats, as the author here addresses. Being limited to generic image formats does add a lot of weight to sites, but also means all browsers can view them.

kolmogorov complexity and mutual information are the things that explain the difference.

The super mario rom is a very compact description of the super mario game, in the context of its environment.

The jpeg environment isn't very well designed as an environment to display a single frame from a flat colored video game.

really? this stupid question on Hacker News? O_O of course the bytes surplus derives from the fact that the image is a raster bitmap, while the game program DRAWS the bitmap, just like the difference between rasterizers and vector graphics. Its like having 1kb monolithic kernel which prints "Hello World" and then goes in while(1); forever, and snapshotting the entire screen and saving it in a .png file..

If you're interested in this stuff, the author also hosts the excellent Compressor Head video series: https://www.youtube.com/playlist?list=PLOU2XLYxmsIJGErt5rrCq...

A large portion of the data is noise nowadays. Cell phone camera pixels are so small that they capture very little signal, yet everything is saved at full resolution. It's digital pollution.

Reply on news.ycombinator.com