在GPU的發展中，畫質的細膩度以及省電的效果一直是兩大討論的主軸。而對於移動設備以及嵌入式系統的繪圖器來說，基於沒有無限能源的原因(只考慮電池的話)，省電似乎是一大考量。嵌入式以及手持式裝置中的GPU對於省電的機制，有一套叫做 Tile Based Rendering 的硬體架構，其設計目的是主要是減少使用記憶體頻寬為著手；而其方法特別之處是利用GPU將輸出切割成多個方格(tiles，bins，grids)，進而針對有改變的方格進行繪圖處理以達到省電效果。
其次，在繪圖速度的限制下，程式設計師也想盡辦法地減少繪圖時的運算量，imposter 的使用即是一個例子。Imposter 是一個軟體的機制，物件首先被繪入貼圖中，往後便利用此貼圖來造成繪製原物件的效果。 imposter 的確大量減少了計算量，但同時他會對記憶體的空間作多一點的使用，在本篇論文中即對這點做討論以及研究。我們討論了 billboard 在tile based rendering硬體架構中對記憶體控制、以及存取是否對 GPU 頻寬流量產生效益已達更佳的省電效果。
縱使先前的研究發現使用人型看板貼圖確實減少指令使用量15%，但記憶體的行為(存取)確曾未被考慮。最後，本論文發現了在砌塊式圖形處理器中使用人型看板確實減少了記憶體的存取量達到3%`。不僅如此，本論文也提出了一個可以幾乎對於記憶體零存取的人型看板的理論。此數值尚未被計算。

The graphics processing units (GPUs) commonly used in mobile devices differ from those used in PCs, owing to cost and power constraints. Some embedded GPUs use tile-based rendering, which divides the screen into regions (i.e., tiles). Rendering each tile to a different core within the GPU simplifies some aspects of the chip’s design.
In graphics terminology, a billboard is a viewer-facing rectangle that creates the illusion of a 3D object, due to the fact that the billboard texture (BT) is a rendered 3D image with transparent areas. An imposter billboard texture (IBT) is a BT that does not precisely match to how the 3D object would appear if it were to be actually rendered.
A previous researcher in our laboratory explored a hardware modification to reduce rendering costs by automatically generating and reusing IBTs. As the GPU renders an object, a BT could be simultaneously created. On later frames, the hardware could reuse that BT as an IBT, instead of rendering the object. This required a method for automatically estimating when the IBT would have too much error to be reused.
Although the previous researcher found a 15% reduction in computation, the memory impact was not measured. This present work measures that impact, and finds a 3% decrease in DRAM accesses. Going further, this thesis also proposes a new approach that achieves benefit without incurring memory costs. The final number for this is not yet determined.

1. Introduction 1
1.1 Graphics Rendering 2
1.2 Imposter billboards 4
1.3 Automatic Imposter billboard textures (IBTs) 5
1.4 Tile-based rendering GPUs 6
2. Background on Tile Based Rendering 13
2.1 The basics of triangle rendering 13
2.2 Tile-based fragment shading 15
2.3 Texture compression - ASTC 19
2.4 The memory structure of a tile-based rendering GPU 21
3. Related work 26
3.1 Impostor Billboarding 26
4. Methodology 39
4.1 Creation and usage on the IBT in tile-based rendering 39
4.2 Calculation on the cost of memory access 41
5. Experiment Setup 54
5.1 The number of texel memory accessing for GPUs to render the benchmark with or without IBT method 54
6. Result 56
6.1 Result from previous work – the instruction savings 56
6.2 Texel accessing comparison in caches 58
7. Conclusion 62
8. Future work – A zero memory cost theory on IBT methodology 63
Reference 67

[1] Iosif Antochi1 , Ben Juurlink1 , Stamatis Vassiliadis1 , and Petri Liuha2, Memory Bandwidth Requirements of Tile-Based Rendering, http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.79.2488&rep=rep1&type=pdf, 2004
[2] Jhe-Yu Liou, Design, Analysis, and Implementation of a Rasterization Engine based on Tile-Based Rendering Architecture in 3D Graphics http://caslab.ee.ncku.edu.tw/dokuwiki/_media/member:elvis:ljy_master_thesis.pdf, 2009
[3] Sheng-Chang Chang and Prof. Steve W. Haga, “A GPU hardware-based method for automatic occlusion detection and optimization for objects and subobjects, ”,National Sun Yet-Sen University, 2012
[4] Chin-Li Lin,” A Method for Automatically Creating and Using billboards to Increase the Speed of Object Rendering”, National Sun Yet-Sen University, 2014
[5] Mali-T860 & Mali-T880, https://www.arm.com/zh/products/multimedia/mali-gpu/high-performance/mali-t860-t880.php, 2012
[6] “Tile base rendering” , https://en.wikipedia.org/wiki/Tiled_rendering
[7] GLBenchmark, http://gfxbench.com/result.jsp
[8] Frame buffers, https://www.open.gl/framebuffers
[9] Billboarding, http://www.lighthouse3d.com/opengl/billboarding/
[10] Kenneth Rohde Christiansen∗, “The use of Imposters in Interactive 3D Graphics Systems” Department of Mathematicsand Computing Science Rijksuniversiteit Groningen Blauwborgje 3 NL-9747 AC Groningen
[11] Tomas Akenine-M¨ oller, Eric Haines, Naty Hoffman. Real-Time Rendering(Third Edition), 1999
[12] The Mali GPU : An Abstract Machine, Part 2 – Tile-based Rendering https://www.community.arm.com/graphics/b/blog/posts/the-mali-gpu-an-abstract-machine-part-2---tile-based-rendering, 2014
[13] 移動GPU三種主流架構優缺點, https://read01.com/O3keBj.html
[14] Midgard GPU Architecture, ARM, https://community.arm.com/graphics/b/blog/posts/mali-midgard-family-performance-counters, the Architecture for the Digital World, 2014
[15] Efficient Rendering with Tile Local Storage, Marius Bjorge, Staff Engineer, http://www.geomerics.com/wp-content/uploads/2014/11/Efficient-Rendering-with-Tile-Local-Storage.pdf, 2014
[16] Technique from ARM, https://www.arm.com/products/graphics-and-multimedia/mali-technologies
[17] Mali Graphic Architecture, https://community.arm.com/graphics/f/discussions/6906/mali-graphic-achitexture?pi69=2
[18] GPUs memory latency, http://www.sisoftware.net/?d=qa&f=gpu_mem_latency
[19] RGB565 format http://www.theimagingsource.com/en_US/support/documentation/icimagingcontrol-class/PixelformatRGB565.htm
[20] ARM Unveils Details of ASTC Texture Compression at HPG Conference
https://community.arm.com/graphics/b/blog/posts/arm-unveils-details-of-astc-texture-compression-at-hpg-conference---part-1, 2014
[21] Adaptive Scalable Texture Compression, C. Dachsbacher, J. Munkberg, and J. Pantaleoni (Editors) J. Nystad1, A. Lassen1, A. Pomianowski2, S. Ellis1 and T. Olson1, 2012
[22] Round robin, http://whatis.techtarget.com/definition/round-robin
[23] Frame Buffer Object, http://www.songho.ca/opengl/gl_fbo.html