Resumen de Exploiting frame coherence in real-time rendering for energy-efficient gpus
The computation capabilities of mobile GPUs have greatly evolved in the last generations, allowing real-time rendering of realistic scenes. However, the desire for processing complex environments clashes with the battery-operated nature of smartphones, for which users expect long operating times per charge and a low-enough temperature to comfortably hold them. Consequently, improving the energy-efficiency of mobile GPUs is paramount to fulfill both performance and low-power goals. The work of the processors from within the GPU and their accesses to off-chip memory are the main sources of energy consumption in graphics workloads. Yet most of this energy is spent in redundant computations, as the frame rate required to produce animations results in a sequence of extremely similar images.
The goal of this thesis is to improve the energy-efficiency of mobile GPUs by designing micro-architectural mechanisms that leverage frame coherence in order to reduce the redundant computations and memory accesses inherent in graphics applications.
First, we focus on reducing redundant color computations. Mobile GPUs typically employ an architecture called Tile-Based Rendering, in which the screen is divided into tiles that are independently rendered in on-chip buffers. It is common that more than 80% of the tiles produce exactly the same output between consecutive frames. We propose Rendering Elimination (RE), a mechanism that accurately determines such occurrences by computing and storing signatures of the inputs of all the tiles in a frame. If the signatures of a tile across consecutive frames are the same, the colors computed in the preceding frame are reused, saving all computations and memory accesses associated to the rendering of the tile. We show that RE vastly outperforms related schemes found in the literature, achieving a reduction of energy consumption of 37% and execution time of 33% with minimal overheads.
Next, we focus on reducing redundant computations of fragments that will eventually not be visible. In real-time rendering, objects are processed in the order they are submitted to the GPU, which usually causes that the results of previously-computed objects are overwritten by new objects that turn occlude them. Consequently, whether or not a particular object will be occluded is not known until the entire scene has been processed. Based on the fact that visibility tends to remain constant across consecutive frames, we propose Early Visibility Resolution (EVR), a mechanism that predicts visibility based on information obtained in the preceding frame. EVR first computes and stores the depth of the farthest visible point after rendering each tile. Whenever a tile is rendered in the following frame, primitives that are farther from the observer than the stored depth are predicted to be occluded, and processed after the ones predicted to be visible. Additionally, this visibility prediction scheme is used to improve Rendering Elimination’s equal tile detection capabilities by not adding primitives predicted to be occluded in the signature. With minor hardware costs, EVR is shown to provide a reduction of energy consumption of 43% and execution time of 39%.
Finally, we focus on reducing computations in tiles with low spatial frequencies. GPUs produce pixel colors by sampling triangles once per pixel and performing computations on each sampling location. However, most screen regions do not include sufficient detail to require high sampling rates, leading to a significant amount of energy wasted computing the same color for neighboring pixels. Given that spatial frequencies are maintained across frames, we propose Dynamic Sampling Rate, a mechanism that analyzes the spatial frequencies of tiles and determines the best sampling rate for them, which is applied in the following frame. Results show that Dynamic Sampling Rate significantly reduces processor activity, yielding energy savings of 40% and execution time reductions of 35%.
