The Design Of Tile-Based Graphics Processing And High Quality Algorithms In3D Rendering

The Tile-Based architecture is one of the mainstreams in embedded graphics system design domain. The most notable feature is that the screen is divided into rectangle areas, so called tiles, in where the rendering processes are executed one by one. Compared with the conventional graphics pipeline, the advantage of Tile-Based architecture is the elimination of bandwidth requirement from depth test. Nowadays the key to good graphics performance and low power consumption is to reduce the internal traffic between the processing elements and the memory, therefore the Tile-Based architecture is suitable for the state of art of embedded graphics system design. The other side, the deferred rendering technology is the nature of Tile-Based architecture which removes the redundancy calculations of texture mapping and reduces the bandwidth further. Today the Tile-Based architecture is used by major manufacturers, such as the series of PowerVR MBX/SGX from Imagination Technologies Ltd and Mali series from ARM et al. but there are some basic disadvantages within Tile-Based architecture. For example the compatibility troubles need to be fixed between Tile-Based architecture and conventional3D API standards such as OpenGL, OpenGL ES, DirectX et al. And the low bandwidth advantage is based on the careful design of intermediate data structure. Accordingly I choose the graphics algorithm design for embedded GPU under Tile-Based architecture to be my key research domain. At the same time some high quality rendering algorithms for embedded are involved. The production of this thesis can be included as follow.1. According to the sequential semantics stream programming model in the conventional3D API standards, we find out the conflicts between Tile-Based architecture and3D API standards.Currently the series of OpenGL/OpenGL ES API standards maintained by Khronos group and DirectX series from Microsoft are most popular in graphics field. They are based on immediate rendering mode and their program models are both stream programming model with data dependency in program order. The data dependency is as the one in CPU ISA. That is to say, the execution of the current API is relied on the result of the former API. For the Tile-Based architecture is proposed for parallel rendering and the result image is independent of the tile rendering sequence, some conflicts are caused. This thesis analyses the OpenGL ES standards with sequential semantics stream programming model to find out the compatibility issues. For examples, the updating problem of depth/stencil buffer when the alpha test is enabled, the framebuffer read operation in application, the texture generation or deletion in Tile-Based architecture and the ability for rendering complex scene et al.2. The design of efficient Tile-Based architecture compatible with OpenGL ESAim at the compatible issues discussed before, an algorithm for Tile-Based architecture is proposed with an efficient immediate data structure. A visible buffer as tile-size is added on die to hold all the visible fragments for deferred rendering technology. This buffer decreases the bandwidth required by fragments read/write operation with system memory.3. The design of high quality anisotropic texture mapping filtering with lookup tableThe single pixel in mobiles’ screen even needs higher rendering quality than PC, so an anisotropic texture mapping filtering is added into our algorithm which uses the elliptical footprint of pixel with Gussian filter. The lookup table method is used for low calculation complexity and high quality which is smaller than3KB. Furthermore a parameter choosing strategy is adopted for the balance between quality and performance which can be flexible for different environments of embedded systems.4. Fast depth-of-field effect image generationFor interactive depth-of-field rendering, two algorithms are proposed here. First, the respective depth weighted coefficients are employed in traditional fast spread algorithm to simulate the different occlusion rules in the scene’s foreground and background areas, after that the two layers are accumulated together for the result image with merge buffer. This method could avoid the leakage from background to focus objects. Second, some different viewpoints in lens are set to generation the color and depth images of the scene, and then the fast spread algorithm with depth weighted coefficients is processed in each respectively, at last accumulate all the images for final image with depth-of-field effects. The later algorithm has better quality in foreground than the former, but higher calculation complexity and resources consumption.