Back to General and Gameplay Programming

Interface and Gameloop - Thread safety

General and Gameplay Programming Programming

Started by gnmgrl May 02, 2014 07:12 PM

2 comments, last by L. Spiro 10 years ago

gnmgrl

699

Author

May 02, 2014 07:12 PM

Well met!

This might be a rather basic question, but I can't seem to find an answer.

To put it simply, I've got two threads - one for the interface and one for the gameloop of my engine.

Most simple example I could think of:

I've got a button. When I press it, I want to move an object. So what an idiot like me would try is:


void spawnButton_Click(System::Object^  sender, System::EventArgs^  e) {
    Control->ModelEditor->MoveModel("object1", 100, 0, 0);
}

Well, that works like 99% of the time. But sometimes, it crashes - when I try to change the position, while the rendering thread is reading it.

My workaround was to send a request to the gameloop, which it would process on its next runthrough in its own thread. That seems to be working, but it blows up the code by a big deal on the log run, since I have to make request-methods for every single method I want to be accessable from the outside of my engine.

I suppose, some other kind of cross-thread-safety is required here - but I couldn't find something that would be taking considerably less code than my approach. Everything regarding multithreading appears to be a huge pain in the butt to implement, so people try to keep thread-communication to a minimum.

What is the usual approach for this kind of high-traffic-multi-thread-communication? Do you got some good papers to link for help? Everything would be helpful.

Have a nice day!

-gnomgrol

DiegoSLTS

2,113

May 02, 2014 08:46 PM

Have you tried with semaphores? You don't need a lot of code to implement them.

SeanMiddleditch

17,596

May 02, 2014 09:14 PM

My workaround was to send a request to the gameloop, which it would process on its next runthrough in its own thread. That seems to be working, but it blows up the code by a big deal on the log run, since I have to make request-methods for every single method I want to be accessable from the outside of my engine.

This is the best model. If you're finding too much repetition, restructure your code to be more generic. You shouldn't need a whole separate method for each possible edit action. You ideally need only a single "change property" method / command, with possibly a few variations for different types of property edits.

Restructure enough and even in single-threaded code this can simplify a lot. For each edit action initiated by the GUI, create a new instance of some EditAction sub-class (ChangePropertyEditAction, CreateObjectEditAction, etc.). Push this into a queue. Execute the queue later. This pattern also becomes very useful when it comes time to implement undo, as each of your action classes can have both an Apply and an Undo method, letting you play the list of actions forwards or backwards to arrive at the desired state.

Not fully compliant code, but the gist would be:

class IEditAction {
public:
  virtual ~IEditAction() = default;

  virtual void Apply(GameState&) = 0;
  virtual void Undo(GameState&) = 0;
};

class ChangePropertyEditAction : public IEditAction {
  string _propertyName;
  string _targetObjectName;
  string _currentValue;
  string _oldValue;

public:
  virtual void Apply(GameState& gs) override {
    gs.get(_targetObjectName)->setProperty(_propertyName, _newValue);
  }

  virtual void Undo(GameState& gs) override {
    gs.get(_targetObjectName)->setProperty(_propertyName, _currentValue);
  }
};

...

void onMove(string object, vec4 oldPosition, vec4 newPosition) {
  actionQueue.push_back(new ChangePropertyEditAction(object, "position", to_string(oldPosition), to_string(newPosition));
}

The code can use templates to avoid the strings, be more robust, actually compile, etc., but that should give you the idea. For more, look at the "command pattern."

Sean Middleditch – Game Systems Engineer – Join my team!