|Project Name||Stars||Downloads||Repos Using This||Packages Using This||Most Recent Commit||Total Releases||Latest Release||Open Issues||License||Language|
|Cocos2d X||16,784||2 months ago||1,596||C++|
|Cocos2d-x is a suite of open-source, cross-platform, game-development tools used by millions of developers all over the world.|
|Magictools||11,097||21 days ago||1||mit|
|:video_game: :pencil: A list of Game Development resources to make magic happen.|
|Learningpixi||3,912||2 years ago||58|
|A step-by-step introduction to making games and interactive media with the Pixi.js rendering engine.|
|Turbulenz_engine||3,512||a year ago||37||mit||TypeScript|
|Turbulenz is a modular 3D and 2D game framework for making HTML5 powered games for browsers, desktops and mobile devices.|
|Cocos2d Js||1,777||2 years ago||280||mit||C++|
|cocos2d-x for JS|
|Imagineengine||1,648||1||3 years ago||11||October 22, 2018||27||other||Swift|
|A project to create a blazingly fast Swift game engine that is a joy to use 🚀|
|Trenchbroom||1,363||17 hours ago||362||gpl-3.0||C++|
|Cross-Platform Level Editor|
|Source code of HexGL, a futuristic HTML5 racing game|
|Qodot Plugin||847||12 days ago||37||mit||GDScript|
|Quake .map support for Godot|
|Oxygine Framework||656||3 years ago||30||other||C++|
|Oxygine is C++ engine and framework for 2D games on iOS, Android, Windows, Linux and Mac|
Since last year HOTS stopped supporting weaker PC configurations and some people were unable to play the game. For this reason I created a utility that improves the performance of the game and you can download it here.
Watch this video to see the result (you must enable CC/subtitles).
Check Reddit Comments for interesting discussions/suggestions.
I didn’t implement all optimizations that I wanted, but I can add them if there is the interest or if the devs could provide me a better access to the renderer since not all optimizations were possible to implement from “outside” (see Devs Appendix).
At this moment, the utility can:
There are many obvious bottlenecks like high poly models and high resolutions, which decreasing directly improves the performance, but degrades also the visual quality. Switching between singlethreaded/multithreaded rendering brings a huge speed up (my laptop gets 15-25% increase) with no visual quality los, but currently this works only on some integrated Intel graphics cards (you can try your luck with any graphics card, it will be probably just slower).
“0.15 Geometry quality” means that the game renders only 15% triangles of the original geometry, which is enormous reduction, while you can still clearly recognize all the characters. This is because HOTS doesn’t optimize the characters geometry for the game, but uses the same characters as for the menu and therefore all the characters have way more polygons than necessary. Fortunately I found a decimation algorithm that can reduce all the geometry with a nice quality on the background thread almost instantly, so it doesn’t even need file caching. However, it’s not perfect, as you can see some holes on Azmodan model.
“0.5 Texture quality” means that the texture mip-map level is biased by 5 levels and also switching to faster point filtering, while “1.0 Texture quality” means there is no mip-map level bias (therefore original texture quality) and it also uses original filtering.
Is this some kind of hacking? Well, technically yes, but it doesn’t go any further than any your video recording program similar to FRAPS. When FRAPS connects to your game, it searches the DirectX library in your running game process and tells the DirectX to capture the last frame of your game and displays the additional fps info. My utility connects the game the same way as FRAPS and then just tells the DirectX to decrease the quality and/or to distribute the work over multiple cores. Because of this process, some antivirus programs might complain (e.g. Symantec warns about potential threat, while AVG and Nod32 are ok with it).
Is this allowed? Modifying software in general is illegal, but this utility does NOT modify any Blizzard software nor any assets, only communicates with the DirectX. The utility doesn’t even give you any game play advantage for cheating.
Will this work with the next game patch? Yes it will. The utility really doesn’t do anything else than communicates with the DirectX library. The DirectX is always the same on all the PCs. The only way to break the utility would be, if the game significantly reworks the renderer, like switching to DX12 or Vulkan.
I don’t want to install any crap! You don’t have to! The whole utility is just 182KB big, you can extract it anywhere you want and if you don’t like it, just delete it! That’s it!
Now the game looks like a crap, but you can also play it on a piece of crap!
Improved FPS by reducing vertex processing
Fixed startup crash caused by missing DLL (thank you very much for sending me the screenshots, I wouldn't fix it without it!)
Multithreaded mode sometimes caused huge memory consumption when you switched to Windows via Alt+Tab
Fixed crash at startup, game exit, device lost event (thank you very much for sending me the debug log files, I wouldn't fix it without it!).
Switching to Windows via Alt+Tab occasionally froze the application, it shouldn't happen anymore.
Improved custom screen resolution setting - no window mode switch is needed, no middle mouse button is needed, just set "width" and "height" in settings.txt in this utility, that's it.
"texture_quality" also downscales the internal terrain texture size (so it should be slightly faster and save more memory - thank you Ahli for the suggestion).
After pressing F11/F12 threading state, the state was forgotten after the game restart, now it's properly saved in settings.txt as all the other options.
Decreased default "texture_quality" from 0.8 to 0.7 and improved guide.
Lowering TerrainTextureHiResCacheSize and TerrainTextureLowResCacheSize can be contraproductive because setting smaller cache means the cache system doesn't have enough space and has to load new tiles and remove old tiles more often as the camera moves, so I highly recommend to use the default values (simply remove the variables and it will be automatically reset). My utility automatically downscales also TerrainTextureSize proportinally with the "texture_quality", so you don't need to set this value neither (however it needs a game restart for the terrain quality to take effect).
If you are changing screen resolution in variables.txt HOTS user file via width and height, this doesn't allow you to set the resolution lower than 1024x768 (at least it didn't work for me). My utility allows you to set 640x480 which is still playable and seems to have the biggest impact on the performance, so try it out (it allows you to go even lower e.g. 320x240 but the text is not readable anymore).
Optimizing this game without any source code is very time consuming and I decided to stop and simply ask for help either from the devs or some other hard core graphics programmers. The game is protected against debuggers and graphics profilers, so I had to write my own graphics analyzer and I discovered some spaces for optimizations without affecting the visual quality:
The game heavily suffers from uploading the data from CPU to GPU, while there are 2 x 128B constant buffers that are doing cca 300-450 uploads every frame, while more than half of these uploads are the very same data! I bet these are some material properties, that are exactly the same between different objects and therefore don't need to be uploaded multiple times, but without modifying the game I couldn't make this more efficient. I bet that the game sorts the draw calls based on the materials, but maybe it could consider also the material sorting based on the constant buffer content and not upload it if the previous material had the same one (this should cut cca 200-300 uploads to GPU per frame).
After switching to the lowest graphics settings, all my tested machines became CPU bounded. The game issues render draw calls via DX11 immediate context from the main thread, which is later blocking everything else. This gives a huge space for multithreaded optimizations. I injected the creation of the immediate context and created a deferred context instead. It might sound crazy, but since both contexts have the very same interface, this perfectly works. Additionally I had to synchronize map/unmap reads + start with write discard if you are doing writes with no overwrite to index/vertex buffer. After this, I implemented some logic that finishes and executes the deferred command lists on the background thread and that’s literally all you have to do to enable multithreaded rendering. However, the cost of the creation of the rendering commands on the deferred/immediate context should be the same, so this doesn’t give you much benefit, so why to even do that? Well simply because you can now move the execution of the command buffer to the rendering thread and better control when to wait for its finishing (immediate context blocks your main thread unpredictably, because you have no control over when the execution starts/ends). Unfortunately because of the poor driver implementations of DX11 deferred contexts the commands creation takes longer than on the immediate context (so far it seems that only the integrated Intel graphics cards do this properly and 1 player with Gtx750m also reported a significant boost). This gave me a next idea that worked on all tested PCs, but for time reasons I didn’t finish it.
To reduce the commands creation cost, I implemented my own super lightweight command buffer. Because HOTS uses only a limited subset of DX11 api and I know the game limits, I could do a lot of presumptions that the native command buffers couldn’t (DX11 function calls add ref count to resources that has cost of a cache miss and can be completly avoided + we know all buffers and how many times we write to them, which means 0 allocations during map/unmap, etc.). This approach was implemented by many game engines during DX11 era. Unfortunately there is a problem that I had to make extra copies of all map/unmap writes, which wasn’t that an issue. The real problem was during the writes with no overwrite to index/vertex buffers – in this case the map/unmap doesn’t allocate a new buffer and just uses the existing one and the game writes into it without providing me the information which part was overwritten and therefore I didn’t know which part I should copy. Since this type of writes is allowed only for index/vertex buffers, I solved this problem by checking the start/end of DrawIndexedInstanced/DrawIndexed calls (obviously only the range of indices and vertices that is drawn needs the copying). However this detection had its own problems and in the end I implemented it only on a paused replay scene and it resulted in cca 20% boost on all tested machines (obviously this depends on how much the machine is CPU bounded). However since this would be way easier to implement directly with the game source code and the DX11 deferred context worked for me already, I decided to stop the development and continue only if people demand it.
The whole implementation took around 3-4 weeks counting writing own performance GPU analyzer and blind guessing since I had no source code and couldn’t even attach the debugger. I think the multithreaded rendering and constant buffer upload reduction is something worth to implement, so if someone is into this stuff, I would be very happy assisting.
Make sure HOTS runs on your machine without the utility. If HOTS doesn't even run on it's own, nothing I can do :(
If HOTS can run on it's own, but the utility doesn't even start, send me a screenshot of the windows error message to my email and I will try to fix it.
If HOTS and the utility starts, but HOTS game crashes afterwards try it again. If it crashes everytime/often, use this special LowSpecHangDetect version and send me the whole content of a log folder from this special version to my email. This version is extremly slow and will look like it freezes because it generates a lot of extra info (there is no sensitive info, just HW description and callstack in a text file that you can read) and I will try to fix it.
For any questions, feel free to contact me at [email protected] or check the Reddit comments.