Been a little busy with various things, including dealing with my girlfriend’s dog having gotten run over. He’s okay but with a broken hip. Somehow, getting a dev log post out in time for the end of November totally slipped my mind. In any case, if you haven’t been keeping up with the video dev logs. Now would be a good time. The latest one is a bit beefy:

With that said, I’d like to dive a bit more deeply into what was involved in getting the save game hitch down to an acceptable level. As of writing this post, There’s still a hitch doing the actual saving, so the job is really only halfway done. However, I did do some work to get screenshot saving down to an acceptable level, and I think that’s a problem that many other games have, so it may be more useful to cover it in detail.

If you just want a decent solution to the problem, then feel free to skip to the end of the article to the section labeled “The Solution” and take a look at the code posted there. For the rest of you interested in hearing my journey to get there, read on!

The Problem

So why are we wanting to save screenshots as part of our save game routine? Well, the answer here is pretty simple, which is that we want the load game menu to show a screenshot of where the player was when they last saved. That way they can easily tell whether they are picking the right file to load:

Unity’s API really only gives you a few functions to handle taking screenshots. The first one and the easiest to use is ScreenCapture.CaptureScreenshot. It’s a simple to use function that takes a screenshot and saves it to a png file. This is great, right?

The problem is that it causes a massive lag spike. The peak of this single frame spike is close to 66 milliseconds. Needless to say this is unacceptable for a game that is targeting under 16.66ms per frame (and don’t even get me started about 8.33ms 120FPS which is apparently so impossible that Unity didn’t see fit to label it in this graph)

A Million and One Ways to Save a Frame

Okay okay, so lets look at some of our other alternatives. We also have CaptureScreenshotAsTexture and CaptureScreenshotIntoRenderTexture. The second one will come back much later, but for now lets take a look at the first one:

That’s an improvement, however we are still exceeding 33ms, which is slow enough to cause a noticable hitch even if the game were locked at 30fps. Additionally, this doesn’t even include the hit from writing out the file, as we only have a texture in memory.

Alright, so we should look into writing out this screenshot into a file. The easiest way to do this is the following:

System.IO.File.WriteAllBytes("test1.png", ImageConversion.EncodeToJPG(ScreenCapture.CaptureScreenshotAsTexture()));

Here’s the performance graph for that:

So, obviously this is even worse than ScreenCapture.CaptureScreenshot but we should probably have expected that because the fine folks at Unity HQ know what they’re doing (perhaps), and because we’re just taking a function designed for putting a screenshot into a Texture2D and then we’re writing that out to disk.

However this isn’t actually the approach that I used, because I was actually unaware of both CaptureScreenshotAsTexture and CaptureScreenshotIntoRenderTexture when I initially went about writing this code. I don’t know if they are new API additions or if I am just a bit slow. Either way they both got ignored for no good reason. They are perfectly fine approaches for certain cases, but with that said I probably wouldn’t have stumbled upon my ultimate solution if I had known about them. Instead, my approach was using Texture.ReadPixels.

ReadPixels requires implementing in a coroutine delayed to WaitForEndOfFrame(). This seems alright as far as performance goes, not quite under the 60FPS threshold here, but an improvement. However, this graph is simply calling ReadPixels on a 3840×2160 framebuffer, without even writing that out to a file. If we take a similar approach to the way we did earlier, then we’d get the following code and the following performance characteristics:

    IEnumerator CoWriteScreenshotTemp(int width, int height)
    {
        yield return new WaitForEndOfFrame();
        if(screenshot==null) screenshot = new Texture2D(width, height, TextureFormat.RGB24, false);
        screenshot.ReadPixels(new Rect(0, 0, width-1, height-1), 0, 0, false);
        screenshot.Apply();
        screenshot_stored = true;
        System.IO.File.WriteAllBytes("test2.jpg", ImageConversion.EncodeToJPG(screenshot));
    }

Damn, so now we’re almost as bad as CaptureScreenshotAsTexture, while having made the code significantly more complicated. However, because we already need to do this ReadPixels process in a Coroutine, perhaps we can do it over multiple frames, so what might happen if we tossed in a few yield return‘s in there and spread that out over a couple frames, and while we’re at it, why don’t we put the file writing on its own frame?

Okay, performance implications are not always obvious in these high-level languages. We have almost an equivalent hitch across two frames now. Still, it seems like there has to be some fertile ground in this direction, even with all the obfuscation of a high-level language. So perhaps the issue is that we’re saving out the file on the main thread. Perhaps if we spin up a thread to write out the file?

Thread _encodeThread;

IEnumerator CoWriteScreenshotTempMultiFrameWithFileWriteThread(int width, int height)
    {
        if(screenshot==null) screenshot = new Texture2D(width, height, TextureFormat.RGB24, false);
        yield return new WaitForEndOfFrame();
        screenshot.ReadPixels(new Rect(0, 0, width/2, height-1), 0, 0, false);
        yield return new WaitForEndOfFrame();
        screenshot.ReadPixels(new Rect(width/2, 0, (width/2)-1, height-1), width/2, 0, false);
        screenshot.Apply();
        yield return 0;
        rawBytes = screenshot.GetRawTextureData();
        QueuedPath = "test4.jpg";
        _encodeThread = new Thread(EncodeAndWriteFileWithArray);
        _encodeThread.Start();
    }

    void EncodeAndWriteFileWithArray()
    {
        byte[] jpgBytes = ImageConversion.EncodeArrayToJPG(rawBytes,
                          UnityEngine.Experimental.Rendering.GraphicsFormat.R8G8B8_UInt,
                          (uint)Screen.width, (uint)Screen.height, 0, 75);
        File.WriteAllBytes(QueuedPath, jpgBytes);
        Globals.updateLoadGameMenu = true;
    }

Well, in spite of the fact that the above graph looks very similar, the scale on the left has changed, and now we are peaking a bit past 33ms rather than all the way to 66ms. So this is definitely an large improvement. However, we’re still not even halfway to 60FPS… One thing we could do pretty easily is to just add in even more yield return new WaitForEndOfFrame()s and read from the framebuffer over even more frames. However, because we are just reading from whatever the most recent frame was, if the scene changes much, this pretty easily can result in problems such as below:

Notice the visible seam down the middle of the frame. The issue is that the framebuffer is changing while we are trying to read from it. The performance of spreading this out over 5 different slices would look like the following:

Admittedly, I was a bit stuck here, but then someone suggested that I take a look into doing a copy GPU side so that I could simply read from the same framebuffer and not have any of these seams. I apologize for not remember who exactly it was who suggested it. In retrospect it seems pretty obvious, but as you might imagine I was pretty lost in the details at this point.

The easiest way to do this would be to use our earlier friend CaptureScreenshotIntoRenderTexture (I told you it’d come back!). As this will do exactly what the doctor ordered, capture the framebuffer into a RenderTexture. Unfortunately (or perhaps fortunately as the case may be), I didn’t know about this function and so I proceeded to dive into several ways of doing this type of thing, ultimately settling on CommandBuffers. (In the future, this will be the most outdated part of this article, as CommandBuffers are part of the legacy graphics pipeline for Unity and will most likely be removed.)

Okay, I’m going to save the code breakdown for a bit later, because otherwise I’d be putting a bunch of code in here that’s marginally different from the final code. But we’ll just say that we set up a CommandBuffer to copy the contents of the screen over to a second RenderTexture (which is really just a GPU-side texture). As part of that process, we have to blit it to a temporary buffer so that we can flip the image vertically. Due to some vagaries in how Unity handles its graphics pipeline, the image gets flipped after it is post-processed but before it is fed into the final display. We come out the other end with a RenderTexture, only it’s actually vertically oriented correctly, unlike CaptureScreenshotIntoRenderTexture, that just leaves it upside down.

This means that, unlike with our earlier staggered read, we can split the read from this RenderTexture across as many frames as we want without having any seams show up. The performance characteristics of doing this across 5 frames looks like the following:

Okay, so another improvement, but there still seems to be a large hitch at the end associated with getting the raw bytes from a Texture2D, and even the 5-frame-staggered readback from the framebuffer into the Texture2D is exceeding our 16ms frame boundary….enter:

The Solution

So, the ideal way to avoid this readback and conversion hitch is…maybe you can guess?

Well, if you didn’t want to guess, I’m gonna tell you anyway; tell the GPU to do the conversion and readback for us in an asynchronous way. The fact is that we will always end up with a bit of a hitch if we are using the CPU to tell the GPU to give us some texture data right now, because that means it has to stop or finish what it’s doing and focus entirely on that task. Instead, we can tell the GPU to give us that texture data “whenever it has a moment”, this is called a asynchronous readback request and can be done in Unity using the AsyncGPUReadback.Request function.

Because this generates garbage, I instead chose to use the variant that uses NativeArray: AsyncGPUReadback.RequestIntoNativeArray. Note that this does mean that we will want to make sure we dispose of the NativeArray when we are done with it. For my purposes, I really only dispose of the array and reallocate it if the resolution changes.

First the performance characteristics of this approach:

Ah, we are finally at something acceptable. You may notice the peak shows slightly above the 16ms line, so it’s clear there is still some hit from the async readback, however in practice the stutter is not noticeable. I may revisit this further before release to see if I can squeeze out a bit more performance here. But for now I am happy enough with this part of the equation. For Taiji, I still have to do some work to improve the actual writing of the save file, but the screenshot taking is “good enough to ship”, as far as I’m concerned.

So, here’s a snippet of something resembling the final code that I used. This is not exactly a plug and play class, but should be a good basis for getting something working in your own projects. I tried to at least have everything necessary in the snippet:

string QueuedPath;
NativeArray<byte> imageBytes;
byte[] rawBytes;
Thread _encodeThread;

public Material screenshotFlipperMat;

Vector2 previousScreenSize;

CommandBuffer cBuff;
public RenderTexture screenRT, tempRT;
    

//Call this function whenever you want to take a screenshot
public void TakeScreenshot(string fileOutPath)
{
    StopAllCoroutines();
    StartCoroutine(SaveScreenshot(fileOutPath, Screen.width, Screen.height));
}

//Helper function that will be called later
void InitializeBuffers()
{
    //We do some checks to see if the resolution changed and we need to recreate our RenderTextures
    bool resChanged = (previousScreenSize.x != Screen.width) || (previousScreenSize.y != Screen.height);
    previousScreenSize = new Vector2(Screen.width, Screen.height);
    
    //The array is Screen width*height*3 because we are going to use a 24bit RGB format (TextureFormat.RGB24)
    if(imageBytes.IsCreated == false) imageBytes = new NativeArray<byte>(Screen.width*Screen.height*3, Allocator.Persistent);
    else if(resChanged)
    {
        imageBytes.Dispose();
        imageBytes = new NativeArray<byte>(Screen.width*Screen.height*3, Allocator.Persistent);
    }
    
    if(tempRT == null || resChanged) tempRT = new RenderTexture(Screen.width, Screen.height, 24);
    if(screenRT == null || resChanged) screenRT = new RenderTexture(Screen.width, Screen.height, 24);
    
    //We build our command buffer, which includes a double blit using a special 
    //material (shader in article) so that we can flip the output from the backbuffer
    
    //This double blit seems to be necessary because CommandBuffer.Blit will not allow us to blit using a material
    //if we are blitting from the backbuffer (BuiltinRenderTextureType.CurrentActive)
    if(cBuff == null || resChanged)
    {
        cBuff = new CommandBuffer();
        cBuff.name = "ScreenshotCapture";
        cBuff.Blit(BuiltinRenderTextureType.CurrentActive, tempRT);
        cBuff.Blit(tempRT, screenRT, screenshotFlipperMat);
    }
    GetComponent<Camera>().AddCommandBuffer(CameraEvent.AfterImageEffects, cBuff);
}


//Function to dispose of our imageBytes from external classes
public void Dispose()
{
    imageBytes.Dispose();
}

//It may be possible to do this in one frame instead of as a coroutine, but I have not tested
IEnumerator SaveScreenshot(string fileOutPath, int width, int height)
{
    InitializeBuffers();
    yield return 0;
    Camera.main.RemoveCommandBuffer(CameraEvent.AfterImageEffects, cBuff);
    QueuedPath = fileOutPath;
    AsyncGPUReadback.RequestIntoNativeArray(ref imageBytes, screenRT, 0, TextureFormat.RGB24, ReadbackCompleted);
}

void ReadbackCompleted(AsyncGPUReadbackRequest request)
{
    if(request.hasError) return; //We just won't write out a screenshot, not a huge deal
    _encodeThread = new Thread(EncodeAndWriteFile);
    _encodeThread.Start();
}

void EncodeAndWriteFile()
{
    rawBytes = imageBytes.ToArray();
    byte[] jpgBytes = ImageConversion.EncodeArrayToJPG(rawBytes, UnityEngine.Experimental.Rendering.GraphicsFormat.R8G8B8_UInt, (uint)Screen.width, (uint)Screen.height, 0, 75);
    File.WriteAllBytes(QueuedPath, jpgBytes);
    Globals.updateLoadGameMenu = true;
}

Additionally, you’ll need a shader to do the backbuffer flipping, so here’s something for that. Better could perhaps be done, but this is based off of an internal Unity copy shader, so maybe not much better:

Shader "Custom/ScreenshotFlipper" {
    Properties{ _MainTex("Texture", any) = "" {} }
    SubShader {
        Pass {
            ZTest Always Cull Off ZWrite Off
 
            CGPROGRAM
            #pragma vertex vert
            #pragma fragment frag
            #pragma target 2.0
 
            #include "UnityCG.cginc"
 
            sampler2D _MainTex;
            uniform float4 _MainTex_ST;
 
            struct appdata_t {
                float4 vertex : POSITION;
                float2 texcoord : TEXCOORD0;
            };
 
            struct v2f {
                float4 vertex : SV_POSITION;
                float2 texcoord : TEXCOORD0;
            };
 
            v2f vert(appdata_t v)
            {
                v2f o;
                o.vertex = UnityObjectToClipPos(v.vertex);
                float2 uv = v.texcoord.xy;
                uv.y = 1-uv.y;
                o.texcoord = TRANSFORM_TEX(uv, _MainTex);
                return o;
            }
 
            fixed4 frag(v2f i) : SV_Target
            {
                return tex2D(_MainTex, i.texcoord);
            }
            ENDCG
 
        }
    }
    Fallback Off
}

Hopefully this has been a fun deep dive for you more technically-minded folks. Lord knows it was a lot of work to write it, but I had fun too. 🙂

Since the last blog post, I finished up revisions on the other areas in the game and sent out a test build to a few people. Still not done getting feedback from that, but overall the game seems to at least be improving. So that’s good news.

Needless to say, there’s quite a number of things that I could write about. But since automated puzzle generation/solving seems to be a pet topic of mine lately for the devlog. I thought I might cover a little bit of how I adapted the automated solver to work for the “snake puzzles.”

There are two main types of panels in the game, if categorized based on how you interact with them. Those are the “toggle” panels, which you click on, and the “snake” panels, which you walk around on. I mostly refer to the snake panels as “walking” panels, but the name still sticks around in the code from the early development days when you weren’t walking on them.

In any case, because of the nature of the input method for the walking puzzles, the same brute force method of finding all possible solutions doesn’t really work.

Admittedly, there is probably a much smarter way of doing things than what I chose to do, but since I don’t want to spend too much time on the auto solver unless I really have to, I chose to simply take the approach of feeding the output of the current solver into a separate “pruning pass.” This works well enough, because all the solutions which are walk-able are really just a subset of the solutions that are possible using the click-to-toggle mechanics.

I chose to essentially take the output of the brute force solver and feed it through two rejection passes. One is a broad (and hopefully fast) rejection pass, which will never reject valid solutions, but will still leave some impossible solutions un-rejected.

The above grid represents a 3×5 solution state that we might be checking to see if it could be done by walking. I didn’t draw any symbols on the panel because it would be irrelevant as that part of solution validation is already done by this point.

So we take this panel and feed it into our first rejection pass. This broad and hopefully fast pass simply checks to see if all of the walked tiles (those would be the ones in white in the above image) are part of the same area. Since as the player can only step on tiles which are adjacent to ones they have already stepped on, this is a guaranteed property of all valid solutions.

Obviously, we can see that there are several different white regions, so this solution is not valid. But the question you might have in your mind is how I would go about that in code? Well, since I already have a function to find all the tiles in a given area, I simply look across the panel until I find a single white tile, call the function to find all the tiles in that area, and then I compare the count of tiles found to the total number of white tiles in the solution state. If they’re equal, then we pass it on to the next pass.

If not, we reject the solution as invalid.

int whiteTileCount = 0;
bool isValidSolution = false;
foreach(bool b in currentState)
{
    if(b) whiteTileCount++;
}

for(int x = 0; x < width; x++)
{
    for(int y = 0; y < height; y++)
    {
        
        if(currentState[x+width*y] && !lockedTiles[x+width*y]) //We only want to look at the area of white tiles
        {
            ArrayList checkedTiles = new ArrayList();
            int num_same_tiles = 0;
            FindAllTilesInArea(x, y, ref checkedTiles, ref num_same_tiles, true);
            
            if(num_same_tiles == whiteTileCount)
            {
                //Could be possible, so we do further testing, we
                //want to OR the values here, because if any spot
                //on the panel allows the solution to be drawn,
                //it's valid
                isValidSolution = isValidSolution | IsSolutionDrawableFrom(x, y, tempState.Clone() as bool[]);
            }
            else
            {
                //We found a disconnected white area, this in and
                //of itself will exclude this solution from being
                //possible on a snake panel
                isValidSolution = false;
                goto EndOfThisSolution;
            }
     }
}

So, we've rejected the previous state and now we will move on to another potential solution state (pictured above). This one passes the first test, as all of the white tiles are connected into one area. So that means it it will get passed on to the second phase.

In this phase, we walk through the panel, tile by tile, from all available starting points (the example panel starts in the bottom left), and we mark off each tile as we go through it. When we reach a branch, we follow out one of the branches until we have eaten all of the tiles. If we do not eat all the tiles by following a branch, that means there is a dead end in the panel and it is not walk-able.

Below is a visualization of this approach.

The code to do this is somewhat more complex than the first pass, and involves a recursive function:

public bool IsSolutionDrawableFrom(int x, int y, bool[] testState)
{
    testState[x+width*y] = false;
    bool returnVal = false;
    
    if(x > 0 && testState[(x-1)+width*y]) returnVal = returnVal | IsSolutionDrawableFrom(x-1, y, testState.Clone() as bool[]);
    if(x  0 && testState[x+width*(y-1)]) returnVal = returnVal | IsSolutionDrawableFrom(x, y-1, testState.Clone() as bool[]);
    if(y < height-1 && testState[x+width*(y+1)]) returnVal = returnVal | IsSolutionDrawableFrom(x, y+1, testState.Clone() as bool[]);

    if(returnVal) return true;

    foreach(bool b in testState)
    {
        //if we reach this point, we've hit a dead end somewhere, this
        //means if find a lit tile that's disconnected from somewhere we
        //could walk to, then this solution must be untenable;
        if(b) return false;
    }
    return true;
}

That's pretty much it for it this week. Perhaps next week I'll discuss a spoilery meta-puzzle thing I've been working on lately. Although if you'd follow the development live on twitch, you'd already know about that!

I promised in the last blog post that I would write a bit about how optimizing the solution-finder paved the way for puzzle generation.

The good news is that I worked on the game a lot in the past week. The bad news is that I still haven’t gotten the offline puzzle generator entirely working. Since it’s a relatively low priority task, and I wanted to ship another quick build to a few people over the weekend, I set the puzzle generator aside in order to make improvements to the starting flow in two of the areas of the game.

Once you have a puzzle solver, making a puzzle generator is actually pretty straightforward. Really all you have to do is plug in random layouts into the solver until you find some that are solvable.

This is already good enough to generate a bunch of solvable puzzles, but it’s a rather blunt tool. It will generate duplicate puzzles, and it doesn’t tell us which puzzles might be worth looking at.

No Dupes

To prevent duplicates, we could generate puzzles in a way that never produces duplicates in the first place. However, this could mean that for any given stretch of time, we could be stuck in a local minima of uninteresting puzzles.

Finding interesting puzzles usually involves a lot of intuition, but generating them could be considered analogous in some ways to a hill-climbing problem. I don’t want to lean much into the analogy, so I’ll just say that since we are hoping to find interesting puzzles that are not very similar to each other, it is good to be as random as possible.

With random puzzles, the only way to prevent duplicates is to keep track of all previously generated layouts. This way, when we want to generate a new layout, we just keep generating layouts until we find one that we have not used already. Since we are already keeping a list of all solvable puzzles, we can easily keep an additional list of all puzzles tested, and check new layouts against the list.

The only real concern here is a trade-off between memory usage and performance. Keeping the list of all tested puzzles in memory will eventually use a lot of memory, but writing the list to disk and parsing that list over and over could be time consuming.

For the time being, I’ve decided to go with the in-memory approach. This simplifies a few things for now, and I can always write things out to disk later.

I chose to use C#’s built-in HashSet collection type to store all the previously generated panels. The benefit of this data structure is that it should be the fastest built-in type for checking if a certain item is on the list (although I have not profiled this).

Currently, puzzles are stored as several arrays. However, for testing the puzzle generator, I’ve just been generating one of these arrays: the symbols array, which stores any symbols that might be on the panel and is just a 1-dimensional integer array of size width*height. (I use a single dimensional array because I’m relying on Unity’s built-in serialization to save the panels and it does not natively support multi-dimensional arrays. I could take the time to write a custom serializer, but I’ve just been accepting the cognitive overhead of doing x+width*y every time I want to do a lookup.)

An annoyance that I immediately encountered when attempting to store generated panels in the HashSet is that when doing the following:

bool done_generating = false;
while(!done_generating)
{
    //symbols is an int[]
    //previousPanels is a HashSet<int[]>
    symbols = GenerateNewPanel();
    if(!previousPanels.Contains(symbols)) 
    {
        //The panel is unique!
        done_generating = true; 
    }
}

previousPanels.Contains() would sometimes let duplicate panels through. That is to say, it would declare that a certain panel state was not in the HashSet when it really should have been.

I would have expected to get hash collision problems, and have panels get rejected as duplicates when they were not, but for some reason the opposite problem arose. I’m not sure why this is. Perhaps there is something funny going on with the implicit call to symbols.GetHashCode() since it’s an integer array (You can find the implementation in the reference source code, but I don’t understand it enough to tell what went wrong). Perhaps I don’t understand hash generation well enough in general.

In any case, I averted my eyes from whatever madness was happening there, and just passed a string in instead and moved along.

…

One other important thing to consider in my specific use case is that puzzles can sometimes be rotated or reflected duplicates. What I mean is that the following three puzzles are identical, as the solution gets rotated and reflected along with the symbols.

Checking for these duplicates is really just a matter of comparing all possible manipulations of a generated panel to the HashSet before declaring it unique.

The Cream of the Crop

Once you’ve got a list of unique puzzles, the problem then becomes how to sort through the many generated puzzles to find ones that are interesting.

This is the thing that I have mostly not finished. I began by keeping track of how many solutions any given puzzle had, as generally speaking, puzzles with only one solution are more interesting to me (although not always more interesting to put in the game, for various peculiar reasons) than ones that have several solutions.

However, I ran into a strange bug where the number of solutions is being incorrectly reported during puzzle generation. I spent a few hours trying to track down some fault in my generation code, or in the way in which I was serializing the finished puzzles to disk.

My best hypothesis at the moment for the location of the bug is that it has to do with the  solution validation code. When checking if a puzzle is solved or not, I am assuming that the panel has been completely initialized. This is usually a safe assumption because panels don’t have symbols moving around at runtime (yet?) and because if I move them around in the editor, the panel gets reinitialized when I’m done.

However, inside the puzzle generator, I’m not re-initializing the entire panel when the symbols change (primarily as an optimization). This may be causing the solution checking to fail because some state has not been properly set.

More on that later. I won’t guarantee that I will be back in the next post to talk more about puzzle generation, as I am beginning to hack away at the tasklist for the next real test build(!)

The list is only 11 items long, but they’re big items.

Header Image from Far Cry Primal reveal trailer.

I’ve been super productive the past week or so. I worked on some fun relatively non-spoilery things, so I can actually talk about them (yay!). Be forewarned; It’s all very technical. But if you’re intrigued by optimization or procedural puzzle generation, press on.

I sent a quick build out to a few people the past couple weeks. Less for playtesting and more for overall thoughts. I’ve still got the next real test build incoming eventually, so if you’re on the list of testers and have been waiting a while….

Don’t worry, I see you.

Anyway, the next build will still be a while, as I have to essentially revise every area in the game. This will mostly amount to cutting lots of puzzles. But after that we can do some testing, and see if I cut too much.

Optimizing Save/Load

One of the “not-testers” this past couple weeks unfortunately lost their progress due to a VM crash. I had assumed this was probably a failure of my saving code (as no untested code is perfect) resulting in a corrupted file, but it turned out that there was no save file at all. The game was only saving when you exited the program. I had forgotten that I disabled periodic saves because there was a noticable hitch if I saved in the middle of gameplay.

Most of this hitch is just because there are so many puzzles floating around in the game world (about 460, with only 230 or so being accessible to the player, and the rest just in the scene as WIP areas or puzzles which I cut for now but that might find a place eventually), but I figured there had to be some performance gains to be had, since I hadn’t really written it in the first place with an eye to performance.

Even still, I tend to avoid doing unnecessary optimization work if I can, so my first inclination was to take the saving and just spread it out over multiple frames. The main issue there is that there is a lot of state that needs to be kept track of to resume saving in an iterative fashion. In theory, this is what coroutines are for, a simple way to make any function you have into an iterative function, by allowing the function to return an iterator which contains its own stack contents. Thus far, I haven’t found a good use for coroutines, because most of the time it’s easier to just store the state in a variable or two and manually resume on the next frame. But aha! Here was the chance for the coroutine to shine!

Except naturally, when I put the function into a coroutine, initializing the coroutine took longer than the actual saving did. So my hitch got worse instead of better. Another “nice” side affect was that I could no longer handle .NETs I/O exceptions because you’re not allowed to use the yield statement (required for a coroutine) inside of a try{} block.

I tried to make the function use asynchronous file operations as well, but this had a similar overhead hitch. Better than the coroutine, but I wasn’t making gains here.

I could’ve continued on. Maybe used threads. Maybe created a coroutine/thread at the startup of the game and just leave it always running along and waiting for saving tasks, but instead I took all this as a hint that I was wasting time on API nonsense and I should actually instrument my code a little and do some real optimizations.

Since the save system is currently human-readable, writing the save files naturally involves writing a lot of text to the file. So, it wasn’t a surprise to see StreamWriter.Write() eating up the majority of the time spent saving. What was a bit more of a surprise was realizing that most of the time spent in that function was on converting integers to strings.

Now, be warned, the following code may disturb you greatly, but this is what I was actually doing for the inner loop which writes the state of all tiles on all panels in the game:

int currentValue;
for(int x = 0; x < (panel.width*panel.height); x++)
{
     if(panel.currentState[x] == true) currentValue = 1;
     else currentValue = 0;
     sw.WriteLine(currentValue); //sw is the StreamWriter we're using
}

Now it should be apparent why I was spending so much time converting integers into strings. For every single tile on every single puzzle in the game, I’m calling the StreamWriter.Write() command and passing it an integer. This means that every time, StreamWriter has to convert it to a string before it can write it to the file.

The reason I was passing an integer into the StreamWriter instead of the booleans directly is because if I simply passed in the booleans, StreamWriter will output them into the file as “True” or “False”, and wanted to write the panel state as “1” and “0”.

So obviously, I made the simple change to the following:

for(int x = 0; x < (panel.width*panel.height); x++)
{
     if(panel.currentState[x] == true)  sw.WriteLine('1');
     else  sw.WriteLine('0');
}

Just this one change got me about a 3x speedup, by avoiding the string allocations and the integer to string conversion. But there’s a few more things I decided to do to make it a bit faster. The first should be pretty obvious:

char[] panelValues = new char[(panel.width*panel.height)+(panel.height)];
int ii = 0;
for(int y = 0; y < panel.height; y++)
{
     for(int x = 0; x < panel.width; x++)
     {
          panelValues[ii++] = (panel.currentState[x+panel.width*y] ? '1':'0');
     }
     panelValues[ii++] = '\n';
}
sw.Write(panelValues);

The main change here is that I simply pre-allocate a char[] buffer for the entire string that represents the state of the panel, and only access the file in one call to StreamWriter. The reason to switch to char[] instead of string should be obvious: strings produce more garbage. And since we’re in a managed language, the less garbage the better.

There’s a few details that are not completely self-explanatory though. One is that I’m now iterating over both width and height. This is just an aesthetic change in that I wanted the panel state to be written 2 dimensionally in the file for better human readability. This is also why the size of the char[] gets and additional panel.height added onto it at the end. I do this because I need an additional column of characters compared to panel states, in order to add the newline characters at the end of each row.

As a bonus optimization on saving, for the situations in which I have no choice but to write integers to the file, I wrote a function to convert integers to char arrays. The function is not very general purpose (obviously), but is much faster than whatever StreamWriter was using.

Note: This code has been updated from the original version due to some bugs pointed out by “Steve” in the comments. If you’d like to view the original buggy code. I’ve posted it here: https://pastebin.com/HNNp7Zug 

    private static char[] IntToCharArray(int input)
    {
        short len = 0;
        if(input < 10) len = 1;
        else if(input < 100) len = 2;
        else if (input < 1000) len = 3;
        else if (input  0)
        {
            output[len-1] = (char)(48+(input%10));
            if(len>1)input = (int)(input/10);
            len--;
        }
        return output;
    }

Although the method of determining how many places there are in the number is pretty dumb looking, it’s super fast. If we wanted to support bigger numbers, we could just add some more lines.

I also did a few optimizations to the loading function. I don’t want to get bogged down too much in the technical details, but essentially I did the same type of thing as I did for saving. I buffer up each line of the panel before parsing it, so as to cut down on string generation as well as calls to the StreamReader.

Procedural Puzzle Generation

So, one other thing that I’ve wanted to work on for a while is improving the tools that I use to make the game. Basically all of the puzzles in the game thus far have been hand designed. However, I’ve made use of a brute-force puzzle solver in some cases.

The puzzle solver worked by using a recursive function. This is a bit complicated, but the important thing to know is that it was very slow. So slow in fact that Unity would kill its own process because I froze the UI for so long.

Since this was a case where I didn’t care too much about the overhead cost of firing up a coroutine, the first thing I did was make the meat of the solver work as a coroutine. That way I could return control to the Unity UI and keep Unity from thinking something had crashed.

I was going to inline the code for the recursive function that iterates over all the possible states of the panel, but it’s too big and unwieldy to be properly readable in this blog’s formatting. So you can look at it on pastebin here: https://pastebin.com/deBHmn0h

A Clever Optimization

One of the biggest performance issues that I was having with the puzzle solver code is that it was an extreme memory hog. This would be bad enough if the code was just executing straight through, but because each recursion of the function was encapsulated in a new coroutine, the memory problem gets out of control pretty quickly. I even ran into a situation (possibly to do with script reloading) where the Unity garbage collector failed to clean up and leaked several gigabytes of heap memory.

The natural solution to this is to change the method by which we traverse all possible states of the panel. Ideally, we would like to change this from a recursive function into a loop, but the way in which we should do that is not entirely clear.

Not to pat myself too much on the back, but when thinking about how to speed this up, I think I made a clever connection. Because the panels are essentially series of bits, we can just think about storing the state of the panel in the bits of an integer. In C# a ulong is a 64-bit integer. This gives us enough space to cover the state a panel with 64 fields in it, or an 8×8 panel.

Importantly, we also can know ahead of time how many possible states there are in a panel. The formula is as follows:

c = 2^w×h

Where c is the count of all possible states, w is width of the panel, and h is height of the panel.

Funnily enough, if we store c in binary, it will take a (w×h)+1 bits.

For example, if we have a 4×4 panel, there are 16 tiles on the panel (w×h). By plugging that into our formula we get 2¹⁶ or 65536 states, which in binary is:

0001 0000 0000 0000 0000

This seems useless, but if we take one less than that, or 65535, we get the following in binary:

1111 1111 1111 1111

Which, if we re-project that as a 4×4 panel where 1 means that tile is on and 0 means that tile is off, is the state in which all tiles are toggled on.

1111
1111
1111
1111

If we then take that number and continually subtract one from it over and over again until we reach 0, we will have covered all possible binary states between. This principle means we can create the following iterative function which will visit all possible states of the panel:

ulong possibilities = (ulong)Mathf.Pow(2, width*height);
while (true)
{
     possibilities--;
     for(int x = 0; x < currentState.Length; x++)
     {
          currentState[x] = ((possibilities & (ulong)(1<<x)) != 0);
     }
     if(possibilities == 0) break;

}

The above code was a bit simplified so that it could be inlined. If you’d like to look at the final complete function, you can find that on pastebin here: https://pastebin.com/5xGi7y5V

In addition to being much more lean on memory, it is much faster. It is also easily resumable with or without coroutines because the state is so simple that we only need to store one variable between runs: possibilities.

TTFN

Wrapping up, we have done some good optimization work on saving and loading. Additionally, because of our puzzle solver optimization, we’ve laid the ground work towards enabling an offline puzzle generator. But I’ll have to cover that in the next devlog post….

I know. I’m cruel… Thanks for reading. 🙂