Semantic Characterization of Shaders Across GPU Workloads: Patterns and Trends

Abstract

As time passes, our 3D graphics applications have become more and more complex. Even though we have thousands of parallel cores in our GPU still it is not sufficient, because our need of 3D graphics has become more realistic and complex too. With the innovation and integration of media objects, 3D graphics application usage has significantly increased. Processing 3D graphics is done by Graphics pipeline. In Graphics pipeline each step, shaders are there to process each frame. First each frame is divided into draw calls and then this draw call executes different shaders and generates 3D image. Now think about it, we need minimum of 16 frames for animated graphics. And this frame is divided into 100 to 1000 draw calls and this draw calls will call 10-100 shader functions for rendering image. So, Most 90 percent of the load of 3D graphics is depends on this shader functions. The goal of this thesis is to minimize shader processing time. So, if we want to improve processing then either we have to develop more capable Hardware or optimize this shader functions. But total shader functions are countless and also hardware capacity and shader optimization has reached to sig- nificant limit. The thesis talks about a way to improve 3D graphics processing and that is, to design hardware for shader group which executes most among all graphics application. The argument behind this is, if we abele to identify that this particular Shader cluster executes 40 percent of total processing, and if we develop hardware for them then execution of shader in hardware becomes much faster and we will get over all approx. 40 percent of improvement. It contains different techniques for identifying such type of shaders clusters, Generating workload, char- acterize the benchmarks, mining traces of workload, clustering most executable shaders, Clustering Co-related Shaders etc...