Rendering Architecture

Cohtml draws the UI (a View) in a user-provided texture.

This happens by calling cohtml::ViewRenderer::SetRenderTarget. The texture must have 4 color channels and depth and stencil buffers attached. On OpenGL, a complete frame buffer must be passed with all its attachments set.

Cohtml performs incremental rendering by updating only the parts of the UI texture that have changed. It will not redraw everything in the UI, which improves performance significantly. The user should not draw anything on the UI texture, except the output of Cohtml. Internally Cohtml uses the Coherent Labs Renoir graphics library.

Cohtml asynchronously records rendering commands that are later executed on the render thread when calling cohtml::ViewRenderer::Paint(). This method draws the new frame if one was recorded, or returns immediately if nothing in the UI has changed.

This is the reason every View object has a corresponding ViewRenderer. The View lives on the UI thread and “controls” the page, while the ViewRenderer is only in charge of drawing it and lives on the render thread.

Backends

On different platforms, Cohtml can draw with different rendering APIs. The API-specific code is encapsulated in a Renoir backend.

The application can also define its rendering backend that uses the graphics facilities it has, or modify the existing backends provided in the Cohtml package.

The backend is registered and used with a call to cohtml::SystemRenderer::RegisterRenderThread. This call tells Cohtml which is the render thread and what backend to use during View rendering.

The backends themselves are objects that implement the renoir::RendererBackend interface. The backends provided in the Cohtml package can be used as a starting point for modification, or for implementing new backends.

Cohtml comes with the following backends:

  • DirectX 11
  • OpenGL 3.3+
  • OpenGL ES 3
  • DirectX 12
  • Vulkan
  • Metal

Rendering Platform Specifics

The Vulkan SDK includes validation layers, which hook into API entry points and provide debugging functionality. In our samples we enable Vulkan validation layers by default on Windows, but not on Android. The motivation for this is, that on Android the layers should be explicitly packaged in the apk. Generally, the following guide should be followed. Alternatively, if you want to package the validation layers in an already built apk, you can use ApkTool and SignApk to do so. This can be achieved by extracting all the tools in the apk output directory and executing a script similar to the following one:

@ECHO OFF

SETLOCAL

SET FileName=ActivityCohtml
SET Platform=armeabi-v7a
SET AndroidNDKPath=C:\Microsoft\AndroidNDK64\android-ndk-r16b
SET ValidationLayerFolder=%AndroidNDKPath%\sources\third_party\vulkan\src\build-android\jniLibs\%Platform%
SET ApkTool=apktool_2.3.3.jar

ECHO %FileName%

java -jar %ApkTool% d %FileName%.apk`

copy /Y %ValidationLayerFolder%\libVkLayer_core_validation.so %FileName%\lib\%Platform%\
copy /Y %ValidationLayerFolder%\libVkLayer_image.so %FileName%\lib\%Platform%\
copy /Y %ValidationLayerFolder%\libVkLayer_object_tracker.so %FileName%\lib\%Platform%\
copy /Y %ValidationLayerFolder%\libVkLayer_parameter_validation.so %FileName%\lib\%Platform%\
copy /Y %ValidationLayerFolder%\libVkLayer_swapchain.so %FileName%\lib\%Platform%\
copy /Y %ValidationLayerFolder%\libVkLayer_threading.so %FileName%\lib\%Platform%\
copy /Y %ValidationLayerFolder%\libVkLayer_unique_objects.so %FileName%\lib\%Platform%\

java -jar %ApkTool% b %FileName%

move %FileName%\dist\%FileName%.apk %FileName%_UNALIGNED.apk

java -jar signapk.jar certificate.pem key.pk8 %FileName%_UNALIGNED.apk %FileName%.apk

del %FileName%_UNALIGNED.apk
rmdir /s /q %FileName%

ENDLOCAL

Advanced Rendering Debugging

Developers can gain further insight into the rendering commands generated during runtime by calling the View::EmitRenderingMetadata method. Metadata markers will be inserted in the rendering command stream that can be hooked to third-party tools like PIX, RenderDoc, Razor etc., or internal engine profilers. The markers will show which elements generate every rendering command in the stream. The backend should also implement the marker-related commands, for the DirectX 11 backend you should enable the DX11_ANNOTATE_EVENTS define, which by default is enabled only in Debug configurations.

Reusing command buffer memory

Prysm uses Command Buffers for serializing the rendering commands, generated by the CoHTML library. These buffers are then passed to the Renoir graphics library, which traverses them and transforms them into Backend Command Buffers, usable by the graphics backend implemented by the client. This transformation process is executed by the Command Processor, which owns the Backend Command Buffers.

Drawing a frame requires allocations for all these buffers so it makes sense to create a pool of them for reuse, and thus decreasing the amount of allocations from the OS.

Clients can control both the amount of objects and size in bytes in the Command Buffer/Processor pools. This is done using the APIs View::QueueSetCacheCountSize, View::QueueSetCacheBytesSize, View::QueueClearCaches and the likes. The new enumeration entries cohtml::InternalCaches::ICACHE_GfxCommandBuffers and cohtml::InternalCaches::ICACHE_GfxCommandProcessors control the Command Buffer/Processor cache counts, respectively.

Font generation on the GPU

SDF on GPU - Font generation on the GPU.

Combating gradient banding

Dithering the gradient rendering - Gradients dithering.

Resource state transitions

Resource state transitions - Support for resource state transitions generated by the rendering library.