Molly Rocket

[casey]

For the last several days I've been working on really annoying things, like full-screen OpenGL failing for no apparent reason on a specific computer (Fujitsu Lifebook), weird vsync behavior, Windows' kernel fucking me every which way on my resource loading thread, etc.

Most recently, I've come across a wonderful little surprise, which is that allowing XInput in my game actually has a rather severe impact on my frame time. It took me a while to figure this out, because it's not something I would normally think of checking... but after blocking it out with my inline profiler and then manually confirming it by commenting out the polling call, I seem to have reached the inescapable conclusion that polling four Microsoft Gamepads takes upwards of around 4-5 million cycles.

The best part? It takes that long even if there aren't any gamepads connected. It even takes that long if the little Wireless receiver is completely unplugged.

How can it take five million cycles to not tell me anything? I mean I'm assuming that it doesn't actually even gather the information during the call, right - presumably the driver gathers the data asynchronously, and the call is just going to make a copy of the current state for your app to process. How can this take a million cycles per controller???

Seriously, can someone from Microsoft send me the secret clue here? Do you have to compute Pi in a bigint or something in order to return the direction of the analog stick or something? What is going on in there?

- Casey

[TomF]

It's not something dumb like it thinks you have ancient analogue joystick attached is it? The sort where you send it a pulse to fill a capacitor, then do a loop until you get a 1 back?

Ah, the days when you'd get a better framerate if you put the stick in the top-left corner. Good times.

[casey]

I wish. It's XInput, so they don't support legacy controllers, so if somehow it was still checking for some that would be pretty ridiculous...

- Casey

[BrianRichardson]

I vaguely remember hearing about similar issues when XInput was polling for controllers that didn't exist or were unconnected. It maybe a timeout inside of XInput. I think the fix was to not poll devices that returned disconnected, or to poll those devices much less.

This is blurry second-hand information from a few months ago, but it might push you in the right direction.

[casey]

I haven't actually gotten around to this to-do item yet. I did briefly test with the transceiver plugged in and not plugged in, but I have not yet tested with controllers in/out yet since I didn't have them handy. I will report the timings when I have them.

- Casey

[sean]

Sounds like this might be someone encountering the same thing ?

[casey]

Awesome.

And XInput was so nice otherwise. I guess they really can't do anything right over there.

- Casey