부산대학교 도서관

Graphics rendering requires various huge amounts of temporary data storages, prohibiting this feature from being implemented on slim embedded devices. To overcome this difficulty, we focus our effort on storage of transparent fragments after rasterization stage. We base our design on the fact that: successive frames typically will have the same or very similar number of transparent fragments located at the same screen pixel location, except in the rare case of scene change. We propose a history based transparent fragment buffer called H-Buffer. Note that transparent fragments arrive in any arbitrary order, making the design challenging. And the storage pressure comes from necessary storage plus internal fragmentation, the latter being resulted from fix-size storage allocation and can be reduced. In this design, transparent fragment counts at all pixel locations are collected for every frame, and be used for storage allocation for the next frame. For the unavoidable case of insufficient storage allocation, our overflow storage allocation assigns neighbor pixel locations to share a given overflow area, in an attempt to reduce internal fragmentation. Easy management and quick access are two major concerns in our design. In evaluation, we used 500 frames from QUAKE4 and DOOM3. Storage requirements are compared against the W-buffer, and the T buffer methods. Compared with the strongest competitor, the T-buffer, results show that our method reduces storage pressure by 25% in QUAKE4 benchmark, and 20% in DOOM3 benchmark. This design idea can be extended to applications where the load change is typically mild and only occasionally abrupt.