RenderMiner: Mining Micro-Architecture Characteristics of Rendering Programs on GPGPU

In recent years, page-table the applications of virtual reality, augmented reality and mixed reality have become more and more popular, all of which are essentially image rendering. Graphics rendering requires a lot of calculations to determine the rendering parameters of each pixel. However, how to reveal the micro-architecture characteristics of GPGPU to help architects increase optimization opportunities and improve the performance of microprocessors becomes imperative. In this article, we propose RenderMiner, a framework which comprehensively characterizes the GPGPU rendering programs at the instruction and memory levels. Specifically, we design instruction mix based on the SASS instruction set, instruction frequency, register reuse distribution and memory call stack to efficiently depict the micro-architecture characteristics of rendering programs and compare with the non-rendering programs. We use RenderMiner to analyze the inherent behaviors of rendering programs combined with machining learning methods through 60 CPU rendering programs, 34 GPU rendering programs, 52 GPU non-rendering programs such as Rodinia and Parboil, and some real-world workloads, such as V-Ray5 Benchmark, Blender, Octane and Kobra Vulkan. After analysis of the rendering programs and comparison with non-rendering programs and the real-world workloads, the key findings of our conclusion are as following: (1) The instruction mix of CPU rendering programs is totally different from that of MiBench, especially at the aspect of arithmetic operation, memory read, memory write and control flow. (2) The instruction-level parallel interval of the CPU rendering program is mainly concentrated from 3 to 6. (3) There is comparable instruction frequency among GPU rendering, Octane, Kobra vulkan and non-rendering programs. (4) The average total instruction count of GPU rendering programs is 355 times more than that of non-rendering programs. (5) The portion of floating point of GPU rendering programs, Blender, V-ray5 benchmark and Kobra vulkan is two times more than that of GPU non-rendering programs.