Kostya's Boring Codec World

As I hinted in my previous posts, I’ve decided to stop working on it. This post should serve as a conclusion, explaining my reasons behind it and mentioning some things about the engine I have not mentioned earlier.

The main factor is the diminishing returns with the rapidly increasing efforts required to get them. I.e. locating the code for loading and parsing different resource files was not so bad (even if a good deal of resources were not parsed immediately after loading but rather treated as some structure data and accessed during use, e.g. model or animation data at each render call). Figuring out the overall engine workflow was not so bad either even if took more time. But things related to 3D are hard because of their non-standard nature (more about it below) and Ghidra not decompiling x87 code correctly in all cases. And the in-world objects interactions are even worse as it is done in the conventional C++ object-oriented fashion (and some bits in less conventional Smalltalk object-oriented fashion) so figuring out what objects are implemented as which classes is not fun, let alone all those variable-length messages that may be sent by them (and handled by a different class).

So I looked at the amount of work before me (implementing GUI, which is easy; implementing 3D rendering stuff, which is too complicated; implementing hardcoded logic, which is too tedious—and that’s before you remember that rooms may also implement their own custom logic in DLL files) and decided that I can stop here as I’m not going to re-implement the engine in any usable form anyway.

Of course I could’ve advanced farther but my inability to make 3D rendering work. I mentioned in the post about room backgrounds how I did not get it exactly right because Ghidra sometimes fails to decompile x87 code introducing variables like extraout_ST1 and you have to guess its value yourself and sometimes outright lying by e.g. using multiplication instead of division—and the x87 code is annoying enough for me to translate it by hoof. Additionally I upgraded Ghidra to 11.0 in a foolish hope that it will improve things, but that was a mistake—not only it did not make any better decompilations for the concerned code but it also made things worse by forcing alignment on function arguments which was not done before (and considering how many functions mixed integers, pointers and doubles as their arguments, I had to correct annoyingly many function prototypes to put them in order again); additionally it changed some interfaces so LX loaders do not work with it yet (it is unrelated but still complicates thing for me; and that is why I’m not so eager to move to the latest version of the software I actively use). In theory spending a bit more time on maths I could get it right but model rendering proved out to be even worse.

I have next to no experience with 3D renderers, especially triangle-based (the books tend to describe ray-tracing approaches instead) yet it is not that hard even for me to understand: surface is split into triangles, each one is defined by its vertices, rotation is performed by multiplying those coordinates by rotation matrix (which is also easy to derive), then you do projection to plane and fill. I’ve managed to make a wireframe model of the object render and after messing with barycentric coordinates I could even make textures appear on some of them.

The problem is that QfG5 engine uses a different approach: normally projected coordinates would be calculated as x/(z/k+1) where k is the distance from the viewer to the projection plane. Inside the engine this +1 bit is missing (yes, I’ve checked the assembly to be sure) which gives unexpected results. What’s worse, even when I export model data into the standard Wavefront .obj format and use a third-party 3D viewer, it fails to apply the textures properly (and sometimes the polygons themselves look very wrong). So it looks like you have to use the engine code—and it hits the same decompilation problems as above (not as bad in this case but it’s still a mix of floating- and fixed-point code with many opportunities for the decompiler to lie).

As for the in-world logic, it does not help that almost everything is hardcoded in the engine. For instance, item IDs are hardcoded and there’s a special table with item properties which contains e.g. sound IDs for viewing or equipping/using the item and message ID3 which tells you which message you should load from 101.QGM e.g. message ID3=18 means that if you load message 1,18,1,1 you’ll get short item description “Basket” while loading message 1,18,5,1 will return “This simple reed basket looks old, but well-cared for.”. Messages with just one different ID are used in many places, e.g. to tell you that your action will fail because of enemy presence instead of e.g. hunger. But the worst of them is a per-character (hero, NPCs or enemies alike) collection of tables, both integer and floating-point ones, that are used for affecting some actions. I did not even bother to find out what they affect exactly.

Well, that’s it. I don’t know what I’ll do next, maybe some small things for NihAV, maybe I’ll look at other Sierra engines to see if they’re more accessible, maybe I’ll do nothing for a while (there are some strategy games that make me waste a lot of time like Battle for Wesnoth, OpenTTD or Settlers II). In either case I had some fun REing the game and learned some things too, I can only hope the next thing I do will be similarly entertaining (and somewhat useful).

This was the last format I haven’t looked too deep at. But as I’m tying loose ends, I decided to revisit it and see what I can get out of it.

As expected, RGD defines the 3D data for the room. I have not figured out all the additional values stored there (and I’m not sure that I got proper regions marked either) but at least the basic understanding proved out to be correct.

So, here’s what I got:

Arcane Island background

Arcane Island RGD

As you can see (especially if you click on the picture to see it in full glory), it is a flat map for the room (or rather like in Doom and Doom 2, a 3D surface consisting of differently raised polygons, pity I haven’t figured out which property is responsible for the heights). Essentially each region consists of a set of edges of different types (I marked them with different colours, you should be able to spot the blue edges around the area and some green polygons around the walkway and columns outlines). There is also a special list of marked regions (yellow ones on the picture but I’m not so sure about that) that are supposed to be walkable and there are even two connectivity matrices for them (for the record, on this particular map there are about nine hundred regions in total and only fifteen are marked as walkable).

From what I saw in the code, this map is used in the game to translate mouse coordinates into destination region, plot path to it, check the enemy line of sight (and projectiles) and so on.

Fun thing is that the format is not compact as the others (as I described it before, it’s a mess of various lists and references to other lists and so on) and so some data is unread. In some cases it is merely 4-8 bytes of padding between different blocks, in other cases it’s way more. For example, in 251.RGD you can see a piece of text “ctile Only. Region 448 contains entry point(s)” plus some other words here and there.

And as I hinted in my previous posts, that’s about it. I intend to write the final post explaining my problems with 3D rendering and why I’m giving up on it. At least I learned a lot more about my favourite game and had some fun while doing it, so it is not time completely wasted.

Since there is not much progress with REing the engine, here I’d like to at least document some formulae from the gameplay mechanics. I realize that most of them are probably well-known already but maybe I’ll cover at least a couple of new ones.
(more…)

Since I have not progressed far in the recent weeks (the current events put me out of mood) I can at least document some bits before I forget them completely. In this post I’d like to outline the overall game engine design.
(more…)

While I’m slowly, very slowly, approach model rendering, here’s something that is between 2D and 3D.

As I mentioned in my previous post, room backgrounds are supposed to be used as a texture on a virtual cylinder. After some investigations I figured out more details and have somewhat working code.

If you forgot geometry as much as I did, here’s a reminder: cylinder is an extruded circle and an image painted on it will be seen in a distorted way with more of the picture seen in the middle (because the distance varies) and sides of the picture being squished (because cylinder curve is at larger angle to the projection plane).

Computing how the pixels should map to the projection is a costly operation so the engine pre-computes tables for the columns to take, actual amount of that column to scale and the scaling step. As the result, during rendering stage all that is needed is to look up the column offsets for each output column, offset inside the column and scaling coefficient (both stored as 16-bit fixed-point for faster calculations). This also explains why background images are stored in the transposed format, it’s definitely easier to manipulate them this way.

Also while fiddling with all this I understood at last what the floating-point numbers in the header mean—they denote start and end angle for the panorama. Those angles are also used in limiting the distance from which the panorama may be seen.

In case of underwater panorama (with its wavy distortion) there’s yet another table in play but I’ve not touched it yet.

The main problem figuring out the code is that Ghidra has problems dealing with x87 code (and it’s hard to blame it, x87 is even worse in some aspects than x86) and I’m not eager to do it by hoof. In the code calculating those projection coefficients step -= delta * 2.0 / width was decompiled as step -= width / (delta * 2.0), I could understand that the called function is arccosine only from the context as Ghidra refused to decompile it and it also failed to recognize the case when common sub-expression was used in two subsequent calculations i.e. instead of y_offset[i] = (int)(offset * 65536.0); scale[i] = (int)((height * 0.5 - offset) * 2.0 / disp_h) - 1; it had scale[i] = (int)((height * 0.5 - extraout_ST0) * 2.0 / disp_h) - 1; where offset=(1.0 - angle * scale) * height * 0.5 but it’s not stored anywhere except in x87 register. As I said, I understand why decompiling it hard but such mistakes require either to try and reconstruct x87 code by hoof or resort to the geometry to derive the proper formulae and I don’t know which one is worse.

In either case here is an example to conclude the post:
(more…)