​​​​Antonios Gogios

The study focuses on interactive 3D rendering pertaining to simulation, CAD software and video game development. It determines and implements optimizations that can be applied to a wide range of scenarios, so it primarily treats effective resource management in an exclusive GPU memory system. It addresses matters like the optimal number and structure of GPU buffers, 3D transformations, reducing overdraw, API draw calls, shader complexity, and pre-computing any data that can be guaranteed to be immutable for the application session.

After a series of tests and measurements to prove its position, the study concludes by suggesting:

​

• Use a frame limiter: By skipping unnecessary frames, much wasted computation is avoided and reserved for useful tasks. Application response time is significantly improved.

​

• Pre-transform immovable objects: By determining and explicitly distinguishing strictly immovable objects in the scene, all unnecessary overhead stemming from their per-frame 3D transformations is eliminated. These objects are instead pre-transformed to World Space either at asset creation time or at load time.

​

• Use collective GPU buffers: To reduce the overhead of re-binding render resources to the pipeline, create collective buffers and address their contents through indexing.

​

• Batch renderable objects and sort them in front-to-back order: Batching will minimize render context switching, while sorting will reduce overdraw.

​

• Use frequency of updates to determine constant buffer count: Constant buffers for DirectX cannot be partly updated, but necessarily update in full. To have the best combination of low update overhead and high flexibility in managing associated data, the data should be grouped into buffers based on its frequency of updates.

​

Also, whenever possible, make use of:

​

• Instancing: Batch-rendering via instancing reduces overall draw calls and associated CPU-GPU synchronization latency.

​

• LODs: If your project constraints allow their use, LOD chains are a worthwhile feature that will significantly lower the scene's polygon count.

​

• Object culling: Employ frustum culling and occlusion culling methods in order to skip rendering of non-visible objects, reduce the number of draw calls issued, and avoid their associated (and effectively wasted) overhead.

​

• Multi-threaded rendering: Modern graphics APIs allow increased multi-threaded flexibility while rendering. On a multi-processor system, much of the rendering process can now effectively take advantage of such concurrency.

​

Some of the above are also mentioned in in my 3D Engine Project features page.

​