Category: programming

So…. it runs! and it runs in places at 60FPS, but in other places….not.

My dell video card has a 64MB VRAM intel GMA chipset, so hardly a gaming laptop, but my aim was to get GTB to at least run on it, even if the framerate sucked. That way, I know people with 128MB or 256MB cards should be fine, and I’d like to keep the min spec as low as I can.

The game already had a ton of stuff you could turn off, such as shadows, shaders etc, but running it on the Dell, and then profiling it using the awesome free intel GMA tools showed up a ton of stuff that I could do to increase the initial 20FPS rate. These were:

• Realise that the refresh rate on the dell was set at 40FPS not 60FPS initally, hence making an artificially low limit. DOH!
• Removing a redundant Clear() at the start of each frame. I fill the screen anyway, so why bother? I don’t use a Z-buffer.
• Removing some render-target sets and clears when the shader options were turned off. With these off, I can render direct to the back buffer, old-school style and save time on render target changes.
• Adding code that detects a jpg when loaded, and mip-maps it. Previously, they had no mip maps at all. Could possibly reduce some memory consumption
• Add a graphical detail slider to options which can turn off a bunch of frilly details like window shadows, and drifting smoke on menu screens.
• Providing a separate list of lower-res textures that get used in some cases when the graphical detail slider is below 25%. Such as mech legs and the shadow maps for scenarios. Any texture of 2048 or higher gets a lower res replacement. I had tried auto-scaling them on load, but this gave unexplained errors and I don’t trust D3DX to do this reliably on all video cards to be honest, so separate low-res textures it is.

I think the biggest wins were the texture-size reductions and the removal of the render target clears. It was interesting to note that the dell considered the game to be GPU limited, despite it being a fairly old and crappy chip (and only a single core). I guess at 1920×1200 res with all the options on for the desktop, things may be very different though.

Things may start to race ahead from here. The game is definitely very playable in its basic form, with the majority of extra work now likely to be the online challenge and integration stuff. That will take months, but still, the end is definitely in sight.

AQTime is my friend. Now… to the optimizemobile!

Well I did it. I am 99.9% sure I finally crushed my random crash bug. Here are the gory details.

It crashed in debug (or release) and seemingly somewhere random. Possibly much later than the actual cause, as is the way with these things. By using Microsoft’s handy CheckHeap functions, I was able to progressively narrow it down to a certain bit of code. (Essentially you check the integrity of the whole heap at two points, then when one of them triggers, you know whether it was inside or outside the enclosed code.

So what I narrowed it down to was some code that command bunkers, hospitals and repair shops run, where they go through their list of ‘supported’ units (the ones they are affecting) and updates them. Essentially this means telling all the units in the list that they now are not supported, then rebuilding a new up-to-date list. For speed reasons this is only done every X frames.

Now the obvious case for a crash is that some unit that was in the list when we added it, has been deleted, and thus when we go to tell them they are no longer being supported, we write to gibberish and cause mayhem. BUT! I know that this is never the case because no units are deleted mid-battle, ever.

Except that it turns out I’m wrong about that :D. They ARE deleted if… I drop a 25 man squad of infantry into a 12-man trench or bunker. At that point 13 of them get deleted. If I dithered when placing them long enough for a nearby command bunker to ‘claim’ them, then when I release them, I set off a time bomb, and after X frames, the command bunker will write to freed memory and hilarity ensues.

All fixed now, obviously. The easy fix would be to not support units until they are placed, but that isn’t nice for the GUI, and also I needed to rewrite the code to use a different system anyway so that multiple bunkers could overlap (they don’t stack though).

2 days. probably 20+ hours of debugging. What a pain. Insert comment about needing to get my act together regarding smart pointers here.

Bah.

My heap is corrupt. No I’m not talking about my car, or my house, but it’s a boring programming thing. Essentially, it means somewhere, I’ve been silly and have written code that writes to memory that it shouldn’t be. This is something that is very very hard to track down.

The trouble with C++, is it lets you do silly things. Not just write to memory that isn’t yours, you could do crazy stuff like define + as meaning subtract, and vice versa. C++ is the programming practical jokers language of choice. There are even competitions to see who can write the most unintelligble crap with it.*

Of course, I try to write the most sensible, intelligible code that I can. After all, I’m the poor schmuck who has to support it, and fend off bug reports from angry customers. The BIG problem with C++ is that when you write some code that does crazy stuff, even when it goes wrong, you might not know about it until much later, maybe seconds or even minutes later, which means hundreds of thousands, if not millions of lines of code have been processed since the actual error.

In other words, you can’t look at a call stack and see where you screwed up, which you can do, in 99% of bugs. Even multithreaded bugs can be at least reasonably interrogated with the call stack and watch window.

With a heap screwup like mine, you are back to pre-debugger, stone age thinking such as.
“When did this start happening”

“What had I done before it crashed? and what had I *not* done?”

The last one is especially helpful because it rules out certain causes. My crash is NOT a result of the level editor, the unit design screen, units exiting the map, nightvision, nighttime, order-giving or path-assigning. It happens in debug, and its only been happening the last week, maximum.

So that means I’ve probabyl screwed up the new pathfinding code or the new AI code. These are two recent areas of much rewriting. I’m combing through them right now. I’ve tried various tools and debug flags to help, but sadly they either don’t run, or slow the game to unplayable speeds, where the bug will likely never trigger.

Fun fun fun!

*C++ is FAST, which is why I use it. If you think 2D games don’t need speed, try cloning GSB or GTB to run at 1920 1200 res and run smoothly on a 4 year old video card. You NEED a lot of speed for all this stuff.