700,000 lines of code, 20 years, and one developer: How Dwarf Fortress is built

[Ed. note: While we take some time to rest up over the holidays and prepare for next year, we are re-publishing our top ten posts for the year. This is our number one post of 2021! Thanks for reading and we’ll see you in the new year. ]

Dwarf Fortress is a type of oddball ardour tasks that’s damaged into Web consciousness. It’s a free sport the place you play both an adventurer or a fortress filled with dwarves in a randomly generated fantasy world. The simulation runs deep, with new video games creating a number of civilizations with histories, mythologies, and artifacts. 

It has turn into infamous, and rightly so. Particular person dwarves have emotional states, favourite gems, and grudges. And all of it takes place in an ASCII interface that appears imposing to newbies, however feels just like the textual content crawl in The Matrix: craftsdwarf, river, legendary megabeast. 

All the sport is product of 1 developer, Tarn Adams, aka Toady One, who has been engaged on Dwarf Fortress since 2002. For the primary 4 years it was an element time challenge, however since 2006 it’s been full time. He writes all of the code himself, though his brother helps out with design and creates tales primarily based on the sport. Up till now, he’s relied on donations to maintain him going, however he’s at present engaged on a model with pixel graphics and a revamped UI that can be obtainable for buy on Steam. 

I reached out to Tarn Adams to see how he’s managed a single, rising codebase over 15+ years, the perils of pathing, and debugging lifeless cats. Our dialog beneath has been edited for readability. If you would like extra, we additionally spoke with Tarn on the podcast.


Q: What programming languages and different applied sciences do you utilize? Principally, what’s your stack? Has that modified over the 15-20 years you’ve been doing this?

A: DF is a few mixture of C and C++, not in some form of normal obeying method, however kind of a multitude that’s accreted over time. I’ve been utilizing Microsoft Visible Studio since MSVC 6, although now I’m on some model of Visible Studio Neighborhood. 

I exploit OpenGL and SDL to deal with the engine issues. We went with these as a result of it was simpler to port them to OSX and Linux, although I nonetheless wasn’t in a position to do this myself after all. I’m unsure if I’d use one thing like Unity or Unreal now if I had the selection since I don’t know easy methods to use both of them. However dealing with your individual engine can also be an actual ache, particularly now that I’m doing one thing past textual content graphics. I exploit FMOD for sound. 

All of this has been fixed over the course of the challenge, besides that SDL bought launched a couple of years in so we may do the ports. On the mechanical facet of the sport, I don’t use lots of outdoors libraries, however I’ve occasional picked up some random quantity gen stuff—I put in a Mersenne Twister a protracted whereas in the past, and most not too long ago I adopted SplitMix64, which was featured in a chat on the final Roguelike Celebration.

Q: What are the challenges in creating a single challenge for thus lengthy? Do you assume that is simpler to do by your self? That’s, since you wrote each line, is it simpler to take care of and alter?

A: It’s simple to neglect stuff! Looking for ‘;’, which is a free methodology however shut sufficient, we’re as much as 711,000 strains, so it’s simply not doable to maintain all of it in my head now. I attempt to identify my variables and objects persistently and memorably, and I depart sufficient comments round to remind myself of what’s happening after I arrive at a spot of code. Typically it takes a number of searches to seek out the precise thread I’m attempting to tug on after I go and revisit some piece of the sport I haven’t touched for a decade, which occurs fairly a bit. I’d say most adjustments are centered solely on sure elements of the sport, so there may be form of an lively molten core that I’ve a significantly better working data of. There are a couple of actually crusty bits that I haven’t checked out since earlier than the primary launch in 2006.

Concerning the relative ease of doing issues on my own, actually for me, who has no expertise engaged on a big multi-person challenge, that is the way in which to go! Folks clearly get good at doing it the opposite method, for instance over within the AAA video games context, and clearly a number of engineers are wanted over there to get issues completed on time. I’d be hesitant to say I can go in and alter stuff quicker than they will, essentially, since I haven’t labored in that context earlier than, but it surely’s true that I don’t have any team-oriented or bureaucratic hurdles to leap via after I need to make an alteration. I can simply go do it. However I additionally should do it alone.

Q: What’s the largest refactor/change that you just needed to make?

A: There have been some refactors which have lasted for months, redoing sure knowledge constructions and so forth, although I’m unsure something is ever a refactor strictly right here since there’s at all times alternatives to push the mechanics ahead concurrently and it is smart to take action when the code data is recent. 

Including the Z coordinate to make the sport mechanically 3D (whereas nonetheless being textual content) was one other one, and actually probably the most mind-numbing factor I’ve in all probability ever completed. Simply weeks and weeks and weeks of taking logic and performance calls that relied on X and Y and seeing how a Z matches in there.

Making the merchandise system polymorphic was in the end a mistake, however that was a giant one. 

Q: Why was this was a mistake?

A: Once you declare a category that’s a form of merchandise, it locks you into that construction rather more tightly than should you simply have member parts. It’s good to have the ability to use digital capabilities and that form of factor, however the tradeoffs are simply an excessive amount of. I began utilizing a “software” merchandise within the hierarchy, which began to get varied performance, and may now assist something from a stepladder to a beehive to a mortar (and pestle, individually, ha ha), and it simply feels extra versatile, and I want each crafted merchandise within the sport have been beneath that umbrella. 

We do lots of procedural era, and if we needed to, say, generate an merchandise that acts partially like one factor and partially like one other, it’s simply method more durable to do this if you end up locked down in a category hierarchy. Including issues like diamond dependencies and all that simply find yourself tying you in knots when there are cleaner methods to do it. If totally different parts can simply be turned on and off, it’s simpler, and lets you do extra. 

I feel some sport builders consult with this as an entity part system, although it’s my understanding that harder-core optimizer folks consider that as one thing else the place you’re really breaking issues down by particular person fields. Utilizing a single object with totally different allotted subobjects is sort of actually worse for cache misses, which is a complete different factor, however the advantages in group, flexibility, and extensibility simply can’t be ignored, and the totally different subfields of the software merchandise aren’t used so typically that it turns into an optimization subject.

Q: Did you run into any points shifting from 32 bit to 64 bit? That seems like a type of issues that was big on the time however has turn into fairly accepted.

A: In no way! I’m struggling to think about a single subject. Thankfully for us, we already had our byte sizes beneath management fairly properly, because it comes up saving and loading the worlds; the format wanted to be nailed down again after we set that up, particularly as a result of we’ve needed to cope with endian stuff between OSes and all that. And we don’t do any gnarly pointer operations or different stuff that may have gotten us in bother. It simply ended up being actually good code for 64 bit conversion because of our different practices, totally accidentally. The primary subject was simply getting the time collectively to make the change, after which it didn’t find yourself taking almost so long as I believed it will.

Q: I’ve seen different video games much like DF die on their pathfinding algorithms.What do you utilize and the way do you retain it environment friendly?

A: Yeah, the bottom algorithm is barely a part of it. We use A*, which is quick after all, but it surely’s not adequate by itself. We are able to’t benefit from a number of the improvements on that (e.g. jump point) since our map adjustments a lot. Typically, folks have used approaches that add varied bigger constructions on high of the map to chop corners, and due to the altering map, these simply take too lengthy to take care of, or are in any other case a trouble. So our method has been to only hold monitor of related parts reachable by strolling. These are fairly simple to replace even when the map adjustments shortly, although it does contain some flood-filling. As an example, if water cuts the fortress in half, it must flood out from one facet and replace a complete half of the fortress to a brand new index, however as soon as that’s completed, it’s good, typically. Then that enables us to chop nearly all failed A* calls from the sport—our brokers simply want to question part numbers, and if the part numbers are the identical, they know the decision will succeed.

It’s quick to take care of, however the draw back is that the part indices are maintained for strolling solely. Which means flying creatures, as an example, don’t have world pathfinding intelligence that’s any totally different from a walker. In fight and some different conditions, we use short-range flood fills with their precise logic to provide them some benefits although. But it surely’s not perfect for them.

I’m unsure we’ll try different constructions right here to make it work any higher. For our map sizes, they’ve all failed, together with some outdoors makes an attempt. In fact, it could be doable with a extremely concerted effort, and I’ve seen different video games which have managed, as an example, some rectangular overlays and so forth that appear promising, however I’m unsure how unstable or massive their maps have been. 

The simplest thought would simply be one thing like including a brand new index for fliers, however that’s a big reminiscence and velocity hit, since we’d want to take care of two indices without delay, and one is dangerous sufficient. Extra particular overlays can monitor their pathing properties (and then you definately path via the overlays as a substitute of the tiles), however they’re laborious and sluggish to take care of because the map adjustments. There are numerous different concepts floating round, like monitoring stairs, or doing a little restricted path caching, and there are in all probability some positive aspects to be made there. We’re actually on the fringe of what we are able to at present assist when it comes to brokers and map complexity, so one thing’ll have to provide if we need to get extra out of it.

Q: On that be aware, you’re simulating lots of issues unexpectedly—how do you handle so many so many actors asynchronously (or do you)?

A: If we’re speaking about asynchronous as in multithreading, then no, we don’t do any of that, apart from the graphical show itself. There’s lots of promise right here, even with microthreading, which the group has helped me out with, however I haven’t had time to dive into. I don’t have any expertise and it’s a bug-prone factor.

Q: Have you ever tried different tasks/applied sciences alongside DF?

A: Certain! The facet challenge folder that’s migrated between computer systems for the final ten years or so has about 90 tasks in it. A few of them lasted for days, some for a number of years. They’re principally different video games, nearly at all times in different genres, however there are additionally a couple of DF helper tasks, like the parable generator prototype. Nothing near seeing the sunshine of day, but it surely’s enjoyable to mess around.

Q: Along with your ~90 facet tasks, have you ever explored some other programming languages? If that’s the case, any favorites?

A: Ha ha, nope! I’m extra of a noodler over on the design facet, fairly than with the tech. I’m certain some issues would actually velocity up the conclusion of my designs although, so I ought to in all probability at the least study some scripting and mess around with threading extra. Folks have even been form sufficient to provide some libraries and issues to assist on the market, but it surely’s simply troublesome to dam facet challenge day out for tech studying when my facet challenge time is for stress-free.

Q: You will have probably the most attention-grabbing launch notes. What’s your favourite bug and what induced it?

A: It’s in all probability boring for me to say, however I simply can’t beat the drunken cat bug. There’ve been a couple of movies made about it by this level. That was the one the place the cats have been exhibiting up lifeless everywhere in the tavern ground, and it turned out they have been ingesting spilled alcohol after they cleaned their paws. One quantity was off within the ingest-while-cleaning code, and it despatched them via all of the signs of alcohol poisoning (which we added after we spruced up venomous creatures.)

If you wish to attempt Dwarf Fortress for your self, you may download it from their website.

Tags: dwarf fortress, solo developer, video games

More Posts