pouët.net

video

parapete: the safe version is only included because it is required by the rules. I don't consider the 1k version unsafe and I doubt that anyone will need to run the safe version.

keops and others: are you using recent drivers? in either case, i will have a look at it soon, promise.

iq: HLSL does have a noise function, but unfortunately no hardware supports it, yet. You can, however, use it for non-realtime texture shaders.

What i do is basically this:
Generate a 1024x1024 heightmap texture using a texture shader and raymarch from the camera to the landscape surface. Approximate the gradient at the intersection point using some tiny differences and use it for lighting. Use the height of the intersection point in combination with a texture to determine if it is going to be a snowy or rocky pixel. Reuse the raymarching to determine if the pixel is in shadow. Important step: apply some dithering to make up for the imprecision of the raymarching. If there was no intersection with the landscape, intersect the skyplane instead. Add gradient to sky. Add exponential fog. Add boring camera movement, music, etc. Besides, the other half of the story is the actual compression to get this below 1k, but I will not bore you with the details :)

trace and others: sure, having better colors, cameras, music, etc. would be sweet, but then something else would have to be sacrificed. What should that be then? Clouds, shadows, fog, compatibility? I think, I have made a pretty good compromise. I do realize that one would probably have to be a 1k/4k coder to appreciate this.

pan: Do you have the directx end-user runtime? (see nfo)

Optimuz, src: Everything is written in x86 assembly and HLSL. I wouldn't want to give up byte-level control when writing a 1k. You might be able to write something almost as small in C, but it will take a lot of hacking and coersion of the compiler, and even then you will probably never make it generate just the right code. Writing the assembly code directly seems much less frustrating to me.

Thanks for the otherwise nice comments :)

mad: I forgot to mention this: when i hit the surface i perform a few iterations of bisection (binary search for the surface) to improve precision. This, however, only helps if you actually hit the surface. Along the edges I will with high probability just step though the surface without even noticing. I dither everything to even out these errors by offsetting each primary ray by a 'random' percentage of the step size.

I actually use multiple octaves of noise, as the noise itself is pretty dull :)

mad: i use multiple octaves of noise for the dithering as well :)
bored? not at all!

Tigrou: I just tried compressing a minimal 1k HLSL framework. It basically does the following:
1. Creates a window
2. Initializes direct3d
3. hides cursor
4. compiles and sets a constant color hlsl shader
5. fills screen with shader until esc is pressed

The resulting sizes were:
Crinkler 1.1: 715 bytes
Custom 1k compressor: 535 bytes

The code is optimized for the 1k compressor, so you might be able to squeeze the crinkler version a bit.

why would you want to draw a square? (hint hint) :)

tomaes and others: you do realize that it loops forever, right? pretty much like every other 1k out there. Too slow, definitely, but too long doesn't really make a lot of sense :)

yes.
The PE header and the decompressor (context mixing based) take up exactly 255 bytes. Another ~90-100 bytes (compressed) are used for importing and the rest is the actual intro.

the archive has been updated with a win7 compatible binary.