Simple program testing possible implementations of graphics API abstraction layer.

Test was run for 3 main architectures, named Cache, Virtual and Proxy, with 3 classes representing implementations for different graphics APIs. Additionally Cache and Proxy are also tested with or without base class Res that would represent common resource class.

Implementations in Cache and Proxy architectures are hold inside Impl union, allowing to save space on each possible implementation at cost of being always the size of the biggest implementation object.

• Cache
  • One cache based on std::unordered_map with unique key and union of possible implementations.
  • Function GetData(U64) returning implementation data based on global variable representing current active API.
  • Class Base holding key to the actual implementation and accessing it's data via GetData(U64) function.
  • Switching API based on iterating over cache, calling destructor on every implementation and initializing with new one.
• Virtual
  • Class Base from which each implementation is derived.
  • Data returned via normal method overloading.
  • Switching API based on deleting old implementation objects and creating new ones.
• Proxy
  • Class Base with implementations as member data and initializing them based on passed current API type.
  • Accessing data via simple function that decides which API is active basing on state of global variable.
  • Switching API based on call to Switch(Type) method with new implementation type passed in.

Each architecture was measured in terms of accessing implementation data and overall tests time. Test algorithm was run 10 times on 10 000 000 objects. Algorithm:

1. Initialize architecture data.
2. Access and add implementation data for all created objects.
3. If this is last architecture go to point 5.
4. Swap current implementations with next architecture and go to point 2.
5. Delete all architecture data.

This test was performed for all present architectures (tests prefixed with first letter of architecture) and measured in terms of access speed (tests named with used API) and overall test speed (tests named G) of 3 mocked implementations of graphics APIs: DirectX (tests named DX), Vulkan (tests named VK) and OpenGL (tests named GL). Additionally for Cache and Proxy in configurations where Res class is present all tests were given sufix R (global measurement includes both configurations). Example: CDXR - Cache DirectX with Resource as base class, VVK - Virtual Vulkan, PG - Proxy Global.

Results of tests were gathered on few available machines with 64 bit Windows 10 and are presented in Results.txt. Conclusions that comes from performed tests are:

• Cache is almost 3 times slower from other architectures.
• Using Res class as base for container costs 6-7% of performance on Cache and Proxy architectures.
• Memory usage of Proxy is slightly higher than Virtual (but can vary if some implementations are heavy on data).
• Cache uses over 3 times more memory than Proxy in current implementation.
• Proxy is almost 2% faster than Virtual.

According to current data, it is possibly better to follow architecture represented by Proxy since it reduces number of times the memory is allocated and slightly speeds up access time on cost of little more space occupied per object. Another pro might be better usage of cache which could speed everything up even more since Proxy avoids trashing CPU instruction pipeline with accessing virtual methods via traversing vtable for each of them.

Copyright (c) 2021 Marek Machliński