Keith Istvan Farkas

Energy-driven statistical sampling: detecting software hotspots

The effect of texture mapping in enhancing the realism of computer-generated images has made the support for real-time texture mapping a critical part of 3D graphics pipelines. However, the texture mapping is one of the major power consumers in 3D graphics pipelines due to the intensive interpolation computation and high memory bandwidth. This power consuming requires an increased emphasis on low-power design for the migration of 3D graphics systems into portable and future user interface devices. In this paper, we present a dynamically adaptive hardware texture mapping system that can perform adaptive texture mapping based on a model of human visual perception which is less sensitive to the details of moving objects. This flexibility may result in significant power savings without noticeable quality degradation. Our work shows that power savings, up to 33.9%, comes from the reduced offchip memory accesses as the result of an adaptive texel interpolation algorithm. Additional power savings, up to 73.8%, comes from using variable clock and supply voltage scaling in the adaptive computing unit.