Category Archives: programming

Speeding up loading times

August 09, 2014 | Filed under: programming

So I released a video yesterday of GSB2 running on multiple monitors, and in that video, a large space battle loads. It takes about 5 seconds. I find those 5 seconds agonizing. I recently spent half a day speeding up the games startup time because launching the game to test 100+ times a day was starting to buy me. Granted, 100×10 seconds is not very long, but I am incredibly impatient. I sometimes launch 2 apps on 2 different monitors and start typing on app 1 whilst windows finishes creating the window for app 2. I often encounter problems where clicking the google homepage and entering a search term means I miss the first few characters as the intel i7 catches up. I’m VERY impatient.

So how did I speed it up? and what tips are there for coding faster loads?

In this case, it was dumbness on my part. Some code that pre-loads ship designs was checking for whether or not lightmaps existed for each ship texture. Some have them, some don’t. It checked by creating the file and loading it in, failing if not found. HAHAHAHA. Very slow, and not needed. Those ships may well not be accessed this game…

So obviously it was tons quicker to not load the file in then, but just to check for the existence of the file on disk (or arguably to do this even later). Cue big speedup. Another speedup was possible because every instance of that item checked for the files existence, so I cached the result in the ship type and only checked for each file once. Much Much Faster.

That’s fine with a game that has few files on disk, but if you need to do this for 2,000 files, it can be real slow just to do those 2,000 file checks. Accessing actual hard drive data as opposed to RAM is painfully slow. With non SSD drives actual physical platters might need to spin. File access is the killer of load times. File access, and decompression.

Ideally you have data on disk that is pretty much in the format it will take in memory, then there is no decompression, just a straight ‘blap’ into RAM. There is a tradeoff here against debuggability, editability and modability. Some games have 2 file formats, a binary version, used when shipped, and a text version that overrides the binary and is only used during development. That can be kinda buggy, and prevents modders accessing the raw assets which might be a bad thing.

Another trick is to use ‘pak’ or resource files. In principle, and often in practice, these are just big zip files. The file access for the entire games resources becomes a single file read, and you effectively load the whole file-access-table for the game in on startup, vastly minimizing file-access slowdowns in cases where there are thousands of tiny files.

In general the golden rule of load times is to only do what you have to. If you can pre-process assets (into binary formats, for example) then do it. If you can defer some actions until that content is really needed, then do it. There is a natural tendency for programmers to have a ‘initapp()’ function and stick everything in there, but you really don’t need to. The player may well not launch the stats screen in this run-through, so don’t load any assets for it, or indeed, any data for it, and don’t waste CPU time initializing systems that might not even be used.

The best advice I could give anyone for speeding up load times would be to just get a profiler. Some are free, most IDEs come with a built-in one (they often suck), and LEARN how to use it. If you never put any time into fixing it, the chances are your load times are laughably inefficient. Most games are.

So here is a thing, you might be interested in if you use STL, if you don’t…well sorry :D

if use use the sort()n function thats built into an STL list, it guarantees to preserve the order of identical objects in the list. if you use the vector version, all bets are off.

Bloody hell.

So if you have a bunch of asteroids with these Z values


And you use a list to sort them, all is good in the world. if you use a vector, those 3 asteroids at 0 are going to Z-fight like crazy things.

the solution?

use stable_sort()

well call me mr-picky but I think I’d be happier if stable_sort() was the default, and we actually renamed sort() to be take_your_chances_and_do_random_crap_sort().

I presume stable_sort is slower… Luckily I’m not sorting asteroids every frame (that would be NUTS), and I only sort things when I have to, so it isn’t mega critical. it led to a bug where the biggest hulk chunks from spaceships did Z-fighting if theyu weighed ewnough to all have a Z-speed of zero, and thus a Z position(relative) of 0, so when other objects spinning away caused a z-sort, their order got scrambled. If you are a non coder and don’t know what Z-fighting is, it’s a flickering effect you get in 3D games where two images seem to be undecided about which one is in front. You often see it on ‘decals’ such as blood splats on the floor or posters on a wall. It’s annoying…

So I was tweeting that this took forever:


It’s just a dialog box for gratuitous space battles 2., why did it take more than twenty minutes to put together? Now… I’ve seen unity, I know it has all these plug-ins that do stuff like this, and that it’s all very user-friendly etc blah blah. But I’m old school. I’m rocking my own custom-written engine, including all the GUI. That gives me huge advantages (mostly speed) and also some disadvantages. The best advantage is there isn’t anything I can’t make the code do.

The pain with this dialog box came in three flavours.

Flavour one was those circular clock-like indicators. In theory, this is really easy, you can just generate a tri-strip of a lot of polygons and draw a curve thats smooth and crisp as you like, as long as you can spare the vertexs. I’m not drawing many, so it’s not an issue. The problem is, when you do that, you get a too-blocky, too un-aliased clunky mess that just doesn’t look ‘right’ when surrounded by lovely aliased everything. I’m not drawing 3D models, so my game has a  nice smooth look to it, and it jarred badly. So I have a sprite of that curve, and I draw a subset of it using a tri-strip arc. It’s a bit fiddly, ant took a while to get right.

Flavour two was the outline of the right-hand part of the window. It’s a bit complex, as it goes in and out and then around the close button and then loops around those circles, and it has to be really slick too, and ironically in this case it looks better drawn as a crisp 1 pixel line, so there is actual hand-crafted code in there to work out all those positions and curve ncie arcs and lines around them.

Flavour three was speed. I like everything in my game to render fast, including GUI. No point in having a fast engine where 95% of the frame is spent drawing a dialog box. That means ensuring that ouline on the dialog is a single draw call with no fuss, that all those tiny animated bits of fluff in the dialog corners and outside the edges are drawn efficiently, that the calculations on that arc outline are as fast as possible, and that the dialog in general; doesn’t use many draw calls.

It’s all horribly, laughably slow really. I probably have a ‘spare render target’ knocking about that I could use to blap this whole dialog to (BTW they resize dependent on the ship, which adds to the complexity), and then only update it when it changed, otherwise just blapping it as a single quad. In practice, the windows various elements update quite a bit.. but I’m sure I could speed up the module icon rendering with runtime aliasing onto spare render targets. I love all this stuff.

But even I know when I’m getting obsessed and need to move on!

Creeping inefficiency

July 23, 2014 | Filed under: programming

Here is why I reckon that triple-A game runs slow on your PC. The real reason :D

Step 1: Geniuses at Intel / AMD / ARM design an unbelievable;e processor capable of a bazillion operations per second. Efficiency 100%

Step 2: Someone writes a compiler that converts C++ into assembly language / processor specific stuff that makes a lot of assumptions and loses a big chunk of efficiency Efficiency 80%

Step 3: A coder like me waltzes in and writes some code that is as optimized as he can possibly manage, but has deadlines etc and knowledge gaps meaning it’s slightly less efficient than optimal Efficiency 75%

Step 4: He then writes it to run on a single core, because the headache of smoothly spreading tasks over all the cores is unbelievable, plus game code doesn’t multithread easily so… Efficiency 25%

Step 5: Because writing a new engine for each game is un-trendy these days, the coder decides to use an off the shelf engine that makes even more assumptions and compromises… Efficiency 15%

Step 6: Coder #2, not knowing the assumptions Coder #1 made when we wrote those handy functions, calls them every frame instead of once… Efficiency 5%

Step 7: The game gets run on a typical desktop PC, with 30 different apps fighting for CPU and RAM, IM clients, P2P stuff, web browsers, email, all that crapware that shipped with the PC, anti-virus scanners, cool desktop widgets that tell you the weather, music streaming as you play… Final Efficiency 3%.


My numbers are wild guesses, but I reckon there is some truth to it all. For inexperienced coders using off the shelf engines probably boosts efficiency. Maybe some engines under some circumstances on some hardware multithreading is more possible. I can’t help[ fantasizing about a PC that absolutely locked everything down in a big way when you launched a fullscreen game. Turned off everything that could possibly use some CPU or RAM and let the game run like an xbox. Maybe that is what steambox will become?

That’s more likely than many programmers learning how to optimize, that’s for sure :(


Here is a big battle on GSB2 running at 1920 1200 res, on a GTX 670, quad core windows 7 PC. This was taken using the visual C++ concurrency visualizer. 3732 is the main game thread. Green is busy, red is idle, light blue is sleeping (end of frame, waiting for flip). Click to enlarge.


1284 seems to be the thread where directx or the nvidia driver does it’s stuff (not sure which).

7596 2692 and 2788 are my additional threads of GSB2 doing processing. Each of those colored bubbles represents one or more tasks that a thread has grabbed and is working through. The big red stretches are obviously gaps I could potentially fill as I find ways to break apart dependencies of tasks and push more of the main thread into the other cores. It’s obviously already been worthwhile, as I reckon I’m currently doubling the framerate (just about) thanks to multithreading. Almost all the grey blobs are transformation of particles within particle emitters, packed into arrays. These are too numerous and cause too much thread-scheduling right now so I might make those arrays bigger, or even dynamic sized.


There are no comments yet