The HardwarePalette will now grow its palette buffer and texture in power-of-2 increments. This avoids it having to allocate memory for a full 256x256 texture up front. In practice the default mods use 22 or 23 palettes so a 32x256 texture is used. This means both the buffer and texture save neatly on memory. Additionally, HardwarePalette.ApplyModifiers sees a nice speedup as it has to transfer a much smaller amount of memory from the buffer to the texture.
To facilitate this change, the MaxPalettes constant is no more. Instead the PaletteReference deals with the calculation of the index and this is passed into the appropriate methods.
The OrderBy overload that takes an int generating key selector is faster than the one that requires a custom comparer. We extract a function from the ScreenZPosition function that determines the Z position of a WPos with an offset, but does not account for the tileset height. For the ordering function this is fine as only the relative magnitude of the comparison keys matter, so we don't need to spend time adjusting for the tileset height, as that won't affect the sort.
To resolve the ambiguity introduced when the introduction of isometric maps meant that cell and map coordinates were no longer equivalent, a new type has been introduced so they can each be represented separately.
Automatically formatted all files via VS. This generally corrects indentation, removes trailing whitespace and corrects misplaced tabs or spaces. Manually tweaked a few files where required.
Avoid allocating memory and resources for graphical elements that will never be drawn when starting a dedicated server. This reduces the server memory footprint by approx 17 MiB.
Additionally exposed an InitializeSettings method on game to initialize the global settings so that other classes can set up all the secret dependencies on the global settings required.
Textures, FrameBuffers and VertexBuffers allocated by the Sdl2 Renderer were only being released via finalizers. This could lead to OpenGL out of memory errors since resources may not be cleaned up in a timely manner. To avoid this, IDisposable has been implemented and transitively applied to classes that use these resources.
As a side-effect some static state is no longer static, particularly in Renderer, in order to facilitate this change and just for nicer design in general.
Also dispose some bitmaps.
By allowing a palette to be copied to an array, a speedup can be gained in HardwarePalette since palettes can be block copied using native magic rather than having to copy them item by item. We transpose the buffer in HardwarePalette in order to allow a contiguous block copy into this buffer, transposing again in the shader to keep everything correct.
The buffers in SequenceProvider can be freed if Preload is called, since we know everything is loaded. A SequenceProvider is created for each TileSet is use so this saves memory for however many tilesets had been used in the game. This will be at least one for the shellmap, and often more.
The MapCache loading thread is kept alive for 5 seconds after it last generated a map (in anticipation of more requests). Once this time expires the thread is allowed to die, as it is unlikely there will be more requests in the short term. At this time it is ideal to force the changes to be committed to the texture so we can release the buffer. As well as marking the buffer for release, we must access the texture to force the changes stored in the buffer to be written to the texture, after which the buffer can be reclaimed.
Additionally, when starting the MapCache loading thread we must ensure the buffer is created from the main thread since it may query the texture object which has thread affinity. After that the buffer may be modified freely on the loading thread until released.
The managed byte buffer is created on demand, meaning a newly allocated sheet will not waste memory holding onto the buffer until some changes are actually required to be written. This avoids a newly allocated sheet wasting memory on buffers that do not differ from their backing texture.
We split the caching SpriteLoader into a SpriteCache and FrameCache. SpriteLoader instead becomes a holder for static loading methods.
Only a few classes loaded sprite frames, and they all use it with a transient cache. By moving this method into a new class, we can lose the now redundant frame cache, saving on memory significantly since the frame data array can be reclaimed by the GC. This saves ~58 MiB on frames and ~4 MiB on the caching dictionary in simple tests.
