high performance computing on graphics processing units: hgpu.org

Optimizing programs for Graphic Processing Unit (GPU) requires thorough knowledge about the values of architectural features for the new computing platform. However, this knowledge is frequently unavailable, e.g., due to insufficient documentation, which is probably a result of the infancy of general purpose computing on the GPU. What makes the modeling of program performance on GPU even more difficult is that the exact value of some "architectural" parameters on the GPU depends on how a GPU program interacts with those features. For example, AMD GPUs show different memory latencies when the memory is accessed with address sequences that have different patterns. Current micro-benchmark suites such as X-Ray are powerless for characterizing the GPU. Clearly, a preliminary for efficient code optimization and automatic tuning on the GPU is a systematic method to measure the architectural features and identify the most basic program characteristics that determine the performance of a program on the new GPU architectures. In this paper, we present a micro-benchmark suite for AMD GPUs that supports the AMD StreamSDK. Our model identifies and measures a series of architectural features and basic program characteristics that are most important and most predictive for program performance on the platform. The features and characteristics include vectorization, burst write latency, texture fetch latency, global read and write latency, ALU/Fetch operation ratio, domain size and register usage for both AMD’s pixel shader and compute shader modes. Our performance model not only generates correct values for those parameters, but also provides a clear picture of program performance on the GPU.