Generics and Boxing Capabilities

Generic classes and Boxing techniques are fairly common practice.  It is often under appreciated.  If we look at a linked list we see that it is using both simultaneously.  It is a great implementation, even though performance can both be improved and hurt by it.  This is a crucial part of understanding how these topics coincide.  The LinkedList<T> from the .NET framework is just one example of this type of concept, and probably in a weaker example.

LinkedList(T) MSDN DOCs

If you want to understand this class further viewing forum topics regarding LinkedList pros and cons would be wise.

This post however is about taking this concept to a new level.  In many application development practices we work with Hierarchies.  They're often coded to be specific to the application.  Sometimes, we implement base classes for this architecture because we use it so often, but how much further can we take such a technology?  Can it be improved upon?  Obviously this is rhetorical, and theoretical.  Not only can we make the assumption there are no code practices that cannot be improved upon, we can also use this case as an example of such.

The run down of how we using Boxing and Generics to take this tech further is actually not that complex.  Linked lists get their benefit from faster insertion and navigating related objects.  Same sort of reason we use Hierarchies often.  One weakness of hierarchies is always starting at the top.  This implementation is two part.  The data is actually stored for the tree in more ways than one.  And it is retrieved in more ways than one as well. 

Storage
We store the hierarchy in two crucial ways.  Since this is a full generic class we have to create boxing for the nodes.  The nodes store information about the level parent and child relations.  So, we can still start from the top of a tree and iterate down the nodes like normal.  There's a new trick though for faster access.  Sometimes, we need to access only data a few levels deep.  Or all items on a particular level only.  This is why the storage is actually storing nodes in arrays for each level as well.  There is no distinction in this storage regarding parents and children, just all nodes on that level.  There is one trick to this that was discovered by accident, but absolutely incredible.  Children can actually be added first.  The other one that was intentional is that an array of values can be converted into a hierarchy.  My test project acftually converts an array of random integers into a hierarchy driven by digits where multiples of ten define the level.  Looks like this:

1000
   1100
      1110
         1111

The way this system is designed such generation is purely algorithmic or interface driven.  Very cool and very powerful.  May sound kinda lame but if you're into "fancy" code... this is right up your alley.

Retrieval
Since we've stored the data in this way, we can do some incredible things.  One example, say we have a tree 5 levels deep, but the upper 4 levels just contain informational data not important to our actual processing.  In most oldschool hierarchies this would still have to iterate all levels to get the bottom.  Not here.  We can pull just level 5 and not even acknowledge the parents even exist.  Pretty cool huh?  (Pretty simple too....)  This structuring also gives a few benefits in speed.  When retrieving nodes where the parent matching a condition(yes seriously) this is very easy and faster than it would normally be.  If we already have a node we just access it's children which prevents us from having to navigate and compare the entire level.  This is normally how we would do it, just felt it important to note this implementation does not exclude that. 

The ways in which the data is accessed is always the fasted option to retrieve the value desired given the information supplied.  In many cases faster than it would normally be from a normal hierarchy setup, but admittedly not all.  It's functional and optimized in many ways, but as many practices teach us, functionality often sacrifices performance.

Code Snippet
Below is a quick snippet from the integer example mentioned:  (There are other parts not visible here excluded to avoid confusion)

No, I am not kidding... It's that simple... And yes... this works flawlessly after 100s of tests.

Fun stuff right?  =P

Extension Method Insanity!!!

Extension methods are something I've known about for a while, but I have realized some new things in their practice recently that make them only that much more fascinating.  The first thing is to understand exactly what they are.  An extension method is "actually" a static method where the first argument designates it being available as an instance method.  Here is a snippet of one for syntax.

The above snippet is a simple extension that can be called on ANY IEnumerable<T> to wrap it within a TSList object(this is just a thread safe list I created).  Pretty cool, but doesn't do much.  This is way I would use them mainly when I first discovered them.  It works well to reduce code lines a good bit.

RECENT DISCOVERY #1

So, given that extension methods are trully static methods, we can actually call them on any field/property that matchs the first argument type.  The value CAN be null!!!  This works well in a few other cases.  Here's another snippet of a simple use case.

So, the above extensions allow any  event handler to be called using one line instead of the normal two to ensure it's not null.  Yawn.... these two examples are really not that exciting.

RECENT DISCOVERY #2

Extensions can actually be used to do incredible things with reinstantiating objects, and thread safety during that change.  In the snippet below I am taking an array, we don't need to know what kind, and resizing it while maintaining thread safety.  There are other extension methods for pulling the values etc. while utilizing the same thread safety techniques.

The using brackets are due to a Smartlock class I use which is instantiated by the CreateLock EXTENSION method on objects with the ICollection interface.  If you review this closely you should see that it pulses both the previous lock and the new lock when a new array is created.  It does this to ensure that the other thread safe extensions will be pulsed if needed one way or another.  (This has been tested for stability with success).

RECENT DISCOVERY #3

Ok, this is where things get a little more interesting and becomes that much more useful with all the above practices considered as well.  Using polymorphing types extension methods can be used to change objects state and type in method chains.  I don't have quite enough space here to really showcase the whole thing, but my posting isn't a tutorial anyway, just something to think about.

When looking at this one note that it doesn't always return the same object.  It may return a "MultiSelector" or the argument passed.  There are others to force return of the parent or a child etc etc.  Methods like the first ToList example make this less fantastic, it just shows a more clear example of how extreme this can go.  If you study up on interfaces and polymorphism you can do amazing things with extension methods.  Like converting Enums to arrays.  That one is really fun =)

Comparing RuneEngine to Unity3D - Test #1

Last night I performed my first official render comparison test with Unity3D's latest released version.  Unity 4.0 may be a significantly different story, but I use what I have access to for now.  I was unable to test against Unity Pro version which supports static instancing which may make a significant difference in the results, but here is my test case and what was found:

Test Case --
20,000 Cubes assigned to Diffuse texture
500 Point lights
0 moving objects
Deferred Rendering configured.
720x480 resolution
Physics Disabled
Input Disabled

Unity3D 3.5.6f4 --
Editor Crashed
Debug build Crashed
Release build Runs
FPS in Debug: n/a
FPS in Release: ~3-4

RuneEngine v2 --
Editor n/a
Debug Build Runs
Release Build Runs
FPS in Debug: not tested due to Unity3D results
FPS in Release: ~135-140 as Active Window ~35-40 as Inactive Window

Further tested to determine threshold for 30FPS with 500 deferred lights found that this can support up to 400,000 diffuse textured cubes in conjunction with 500 deferred lights non-moving while remaining above 30FPS.

Ok, a lot of abstract information.  Well, for those who are not aware 28+ is a solid framerate to avoid jerkiness, but some video formats only push 24/25 FPS and remains unnoticed.  Therefore RuneEngine is fully capable of pushing non-jerky frames in this test case and well beyond the specific test case while Unity is not.

My analysis from these results is as follows, the new Update Systems and the Rendering Pipeline are optimal, though are still receiving some improvements.  However, Unity3d is likely to have some overhead that RuneEngine at this time has not accounted for.  Such as the Quad tree sorting.  In the basic scene all objects are rendered, but Unity would still be processing this sort.  Depending on the nature of how they do this that could be the cause of what is seen.  However, I must note that prior to the new Update System in RuneEngine my framerate results were rather similar to what is seen in Unity3D suggesting that this optimization may be the sole reason for the substantial performance difference.

Time will tell if this landslide performance difference continues to be so, but at the moment RuneEngine outperforms Unity3D by ~45-46 times...

I intend to do some further testing with this as time progresses, as well as confirming the performance of Unity3D Professional in comparison.  For now this is a pretty big success for RuneEngine I believe.

RuneEngine V2's Flow System

I thought it would be kind of me to write a larger post about what this system is, how it will be used, and why it is a crucial part of the new RuneEngine V2 feature list.

If you're at all familiar with DirectShow basically imagine that, but re purposed for something... different.  For those who don't know, essentially we're working in Blocks that I call the IFlowBlock interface.  A block defines a procedure that handles a certain type of data.  This part is somewhat different from what DirectShow does.  The data is generic, where in DirectShow it's always a type of media passed by a buffer of bytes.  RuneEngine's Flow system can pass more or less whatever you see fit.  There are just a few rules. So, in DirectShow a sample or data is passed downstream all the way to the end, in every case.  Well... Flow is not going to do the same necessarily.  It can, but generally speaking the end block is going to be a renderer or finalizing state, so it should only be hit when the rest is complete.  Flow data starts upstream and passes downstream until it hits a breaking condition.  Data can be flagged for passthrough transit meaning, it's going all the way to the end or to a certain point before processing itself.

I could go on and on about how data moves and the different possibilities, but the key thing to understand about Flow vs DirectShow.  DirectShow has rules, while Flow is more user defined.  The both have Pins and "Filters" or "Blocks", even "Graphs", but ultimately they are completely different beasts.

Flow is not designed for processing video/image data, it is designed for processing some data in a specified sequence using a system of interchangeable blocks.  Within the upcoming particle system this will be used by the ParticleController.

The blocks though will control a state of a particle.  A particle graph may look rather complicated, but it houses some insane possibilities compared to other particle systems.  And it does this without creating 20 emitters and controllers.  Below is an example of what a simple particle system graph may look like:

This may not be the final example, but in this the idea is despite the appearance, when in one of the velocity blocks the object actually does not receive wind, it just uses velocity to update.  Wind control happens after the velocity state is complete.  So, for a certain time, a particle will be moving in a specified velocity, when reaching it's final output, it moves to wind control.  Particles internally hold a velocity value, so the old velocity should be there still.  Wind control is going to manipulate this.  So, you would see the particle slow down and react to the wind.  Imagine and explosion where the particles move out quickly then get carried away in the wind after.  That is what this somewhat the idea of what this effect would accomplish.

This demonstrates that this system is more sequence driven.  One block does not start til the following is complete, though that is not a forced behavior.  It was the design.  However, particles are small instances so the overall will almost seem like streaming update, it's not exactly.  Now, in theory could you create a loop back effect to handling particles, without changing the underlying system?  Absolutely!!! You could easily create a block that has a loop back output pin which send's items back to another block if the parameter to pass forward has not been met.

Given that this system is newish, and purpose is different than similar systems these are only speculations of ideas.  The actual implementation and design of graphs will come after some testing.  If I find some cool tricks to do with the FlowGraphs they will likely get shared as examples. =)

That's all I got for now though.

New API objects for RuneEngine V2

Ok, the title of this post is almost misleading, this is more about the correction of some features that were slightly flawed in their design which have now been corrected.  So I'm going to talk a little about it.

RuneEngine V2 has had a feature since the beginning referred to as a content reference system.  The idea and purpose of this system is to be able to dynamically load content and handle such implementation as transparently as possible.  Sadly, it only truly came to my attention last night it was not serving it's full purpose.  It wasn't exactly wrong.. but it was defeating it's purpose in existing with the existing implementation.

Now some MAJOR changes were made, everything from the outside looks almost the same, except it is a little easier to use now.  I should be taking that another notch further, but that's a completely different talk.  So, what was changed?  Here's a basic list from the far back perspective:

Removed--
IMaterialReference
IPostShaderReference
MaterialReferenceObject
PostShaderReferenceObject
MultimapTextureReferenceObject (rebuilt to interact with the TextureReferenceObject instead of separate)

Added--
IShaderReference (new material reference, but also serves as post shader reference)
RenderTargetReferenceObject( it was missing )
IContentContainer (this guy... does magic)
ShaderReferenceObject

Ok, that's about it.  As you can see it's a pretty even exchange, but the added were completely missing or are more of a rename.  The multimap texture reference and texture reference objects were completely unnecessary to be separate, I may have to do some wiggling to make it work, but it is done.  Material references and post shading references were extremely redundant hence the spark to make these changes.  In the process is how I found the issues.

IContentContainer was added to resolve the biggest of the issues as well as add a few functionalities that really make this feature sing.  The issue was, the reference objects were calling load methods on instantiation in most cases, and if they didn't.  The components were.  Why not use a normal load call if I'm not going to use these for what they were added for right?  So, first I stripped all calls that forced loading of the content references, then looked closely at what was missing.  IContentContainer was what I came up with.  This interface defines a collection of callbacks that the IContentReference objects call when they are loaded or unloaded.  It has a method for making sure the reference is associated.  That sort of thing.

One interface all it really took?  Not a chance! Now, the IContentReference interface also has some added methods for working with the IContentContainer.  Also, RuneContentManager was extended and improved to work more closely with this new interface as well.

So, rather than being super abstract about what happens let me just explain a few scenarios and how these changes have impacted them.

Scenario #1:
I'm building a game where the scene is rather simple, but the Content takes a long time to load.  I do not want to make a super long load time when a level is loaded.  But, I'm afraid to preload the content bogging down the CPU in the main menu.  In this case I can preload the Scene files but not the content.  Once the level is selected we can prioritize required content loading it first allowing the level to start with partial content and potentially even a partial scene.  The reference objects allow this separation of load times, but not in a way that is forced by the engine.  You can load it whenever you please.

Scenario #2:
I have a large collection of global resources that need to be loaded.  I can put them in a global RuneContentManager and link it to my SceneGraphs allow the components to search through it's content as well.  I can load all these global content items at startup, background or load screen.  And pull them using the RuneContentManager linking systems.  The amazing part here is that the contentReferences will manage unloading and reloading of those assets individually if they're not really being used.  Also, the ContentReferences will pull more directly from the contentManagers now than before.

Scenario #3:
I'm building an Elder Scrolls Clone.... GG!  The newest implementations actually allow full user control over when content is loaded and the IContentReference objects do all the tricky stuff.  You can load a resource in anyway RuneEngine allows and the components and the IContentReferences will receive that callback letting all places that need to know, know that the resource is available and can be used now.  No periodic or frame by frame update required.  All of this leading to a fully seamless platform capability in a very simple way at that.

TextureContentComponent.AssetName = "assetName";
RuneContentManager.Load( "assetName" );

Done.. This like much of RuneEngine, is order independent.  RuneEngine is unique in this element that I've spent a lot of time working around pre-requisite type issues.  The game will not fail when resources are missing it'll just respond to it internally or with user defined code.  Components can be added in any order.  SceneNodes can be activated before or after components are added, it doesn't matter.  The only real prequisite to any of RuneEngine is this line:

RuneManagementService.Startup();

I'm extremely enthusiastic lately watching all of these architectural designs in the engine being realized. So pardon me if I seem ranty.

Anyways, happy coding.

Thoughts after the RuneEngine V2 Code Review

Last night I spent a lot of time going through the RuneEngine V2 source code, updating documentation and reviewing the source for poor organization and garbage code left behind.  This is a good practice to act on with some regularity, but isn't always necessary.

Going through this I had quite a few thoughts.  Absolutely none of them involved considering change.  I found that for 210 code files currently they all seem to be in pretty good order and have great readability.  But that is aside from the point.  The real point is an issue of mentality gained from reviewing the source.

I have spent a lot of time over the past few months procrastinating about not working on the engine.  I am getting heavy back in development with it so I decided a cleanup pass was in my best interest.  I had been procrastinating because of "how much is left."  Well, this review showed me quite a it in the direction of how inaccurate that really is.  There are still plenty of unfinished areas and new features that need implementation, but reminding myself that there are systems and features already implemented that push into professional grade quality affected my opinion of where I am at.  It would appear that by the end of the week I could have any loose ends taken care of and begin development of new features next week.  By new features I specifically mean building the editor.

Now, I've spoken here about the goals of the engine as far as features, but here's a list of what I found that gives me such a confident outlook on the status.

- Deferred Rendering
- Post Processing System
- Dynamic Content system
- Custom material format
- User defined geometry system capable of working to the extent of a simple modeling utility
- Extendable file IO system
- Extendable XNA content resource extensibility (This is small, but huge at the same time, specifically it is    built that shaders with custom structs used can be applied via the material file format with some extension to this area)
- Input System, with plugin capabilities
- Updateless and per-frame draw call removed system (huge optimization)
- Custom SpriteBatch which renders into a deferred scene with textures that support the multiple channels.
- Architecture in a way that almost every part of the engine is 100% self reliant.
- Low-Level and High-Level implementations for most features

As a foundation, RuneEngine V2 has all the necessary pieces to be something really impressive.  Nearly the entire engine is designed for extendability.  This design is going to make moving forward significantly easier than it was in past iterations.

Very soon, after the loose ends are tied up I will be making a post here and on the RuneEngine V2 facebook page.  This post is going to be a demonstration of development with RuneEngine V2.  I want to showcase, how some of this stuff is compartmentalized and how, as a coder using it, things are going to be done.  In many cases a test project only consists of a few RuneEngine lines of code.  If it's much more then it is due to setting up parameters of objects for a less "defaulted" test.

Anyways, happy coding.

My new Development Pattern and why it works

I’m generally not one in the past who has studied other’s development patterns, so I generally just go with whatever works best for me.  I found not real long ago, that doesn’t really work… So, I did actually put some thought into how I worked on my projects.  This has been heavily exercised in the restructure of RuneEngine V2 over the past year.  The code base is the most organized and the most stable, but also the largest it has ever been.  Therefore, I am confident this is working well for me.  So, I decided to share some details in my new design process towards development.

Ok, I’ve always done this, but I’ve taken it more seriously.  Whether I write it down or just visualize it, or even go ahead and set it up in Visual studio.  The first thing I do is define objects that I need in a given area.  I also ensure that I know an approximation of what members I need.  Like I said, I’ve always done this; this is a pretty standard starting point for planning phases.  

The next part is where I determine which methods are used commonly across the objects and/or if there are methods/properties that may come in handy across other parts of the project.  Then, I place all of these common methods and properties in groups and apply them to interfaces accordingly.  Ideally, the first code I write is at this point.  If you’re not overly fond of interfaces and/or believe that they aren’t a necessary part of development you may want to re-evaluate that statement and review MSDN and observe the inheritance trees of .NET objects.  I would agree they’re not required, but when you practice this way you get a few amazing advantages you may not have thought of.

First of all, one example is extension methods for interfaces.  These are a great addition to an interface driven design pattern.  You can define behaviors with these interfaces and avoid rewriting the same logic in every implemented object.  This is more appropriate than inheritance; because it is flat there is no call stack for this.  So, this does become an optimization later.  One major con is that interface casting is expensive, you can control how much time you spend casting interfaces in many ways, so this is not overly an issue.  An example of a major pro would be extensibility of your code base.  Along with the extension methods, other methods that are geared to accept the interfaces and not objects can act on any object that implements them.  This gives phenomenal ability to add whatever you need to adhering to the rules in the interface, and all of your code base will support it.  This is my purpose in designing this way.  One issue that you run into that is not really a pro or a con is that you do have to define everything, every time.  So, it does take more work to add new objects than deep inheritance trees and abstract objects, however this does promote flatter code overall, so again it runs faster.

As I am designing the interfaces, I put a lot of thought into a few things.  Will it return an error code or throw exception?  Do these methods require specific implementation that is not optional?  Can I confine them into fewer methods?  Should this be publically accessed?  These are good questions to ask because they guide the process into making useful interfaces that are not too tedious and remain useful.  It is easy to make an interface that is a total waste.  The most useful interfaces in RuneEngine V2 are the IContentReference and derived interfaces.  The reason these are so useful is because they define rules that allow the RenderManager to use them without the need of any further information about the content.  IMeshContentReference for example, is used for custom geometry, models, animated models, and will be used for terrain as well.  One interface can handle all types of rendering.  To make it even more useful, this interface can be defined by a 3^rd party and game specific code, and guess what?  Now, that custom rendering logic is available to the RenderManager as well.  

After I do all of this and start rolling with the project and ensuring that my interfaces work in their desired environments and the objects I’m building can get their full behavior from this set up.  I go through again and make sure the interfaces don’t have any methods that across the board aren’t used.  The IMeshContentReference did have a few methods that weren’t being used by the RenderManager or the other objects so they were removed.  The objects that relied on them kept these methods, but the interface does not define them so they became object specific.

The last big key to the puzzle is actually one of the more important, that is done from the beginning, but mentioned last for its importance.  CHUNK IT, do not try to list out every interface and all of its methods from the beginning.  Build in chunks.  This gives you time to focus on each individual area closely.  It also allows you to overtime realize a more appropriate location for existing and new interfaces/classes.  I avoid moving things in too big of a hurry, but I’d rather do it now than later.  The longer you wait the more work it will require to do.  I generally define a chunk as a new feature or system.  Anytime I create a new DLL I tried to handle its major features right off the bat. 

A good example of how I work in chunks is this.  When I was starting the restructure of the content system, I knew I wanted to work in ContentReferences as before, but they were going to become interfaces.  So, that was an immediate.  I also, knew already from prior that the ContentReferences did require some uniqueness so I created the Mesh, Material and Texture reference interfaces.  These were the most important and I believed there was a chance I could confine all my rendering code to use them exclusively (I was right).  I didn’t start working on the implementation of any of the reference object implementations until after I had developed them into the RenderManager and the RuneContentManager objects.  I had to confirm their use and entry, what methods they would require before every implementing a single object.  I did find in implementing the objects that the methods required a little tweaking, but there was little to no change during this stage.

The more general stance on this is as follows, start with a list.  Use the list to define rules in the form of an interface.  Start the implementation at the most system level objects that use the interfaces as parameters.  Ensure the interfaces will work then implement the objects.  Make minor changes if necessary as the objects are implemented.  Only work in chunks of implementation as opposed to the entire code base at once.  If an object doesn’t apply in any of these systems and uses no methods that would be required in other objects, there is no purpose for an interface.  Basically, if you define IContentShaderReference when ContentShaderReference is the one and only object that will use it.  Well get rid of the interface.  IContentReference should be enough.  However, if there is any chance IContentShaderReference could be used in another case.  Yes I would include it.  Only exclude an interface definition if you know there is no chance another object will exist in need of those methods that will be passed into the same methods this object will.  Simpler put, if this object could be singing “I’m the only one” Then you probably don’t need an interface.

Most of this may seem pretty general, but as usual this may help you.  It’s just an approach I’ve been using.  You may want to review your approach and see if you do any of this and whether or not the ideas may benefit you.  If you ever find yourself cussing because you don’t have enough flexibility or you have to write large extensions to your entire code base every time a new object is added.  Then this is a good approach for you.

Happy coding.