Towards Interactive Visual Exploration of Parallel Programs using a Domain-Specific Language

Abstract

The use of GPUs and the massively parallel computing paradigm have become wide-spread. We describe a framework for the interactive visualization and visual analysis of the run-time behavior of massively parallel programs, especially OpenCL kernels. This facilitates understanding a program's function and structure, finding the causes of possible slowdowns, locating program bugs, and interactively exploring and visually comparing different code variants in order to improve performance and correctness. Our approach enables very specific, user-centered analysis, both in terms of the recording of the run-time behavior and the visualization itself. Instead of having to manually write instrumented code to record data, simple code annotations tell the source-to-source compiler which code instrumentation to generate automatically. The visualization part of our framework then enables the interactive analysis of kernel run-time behavior in a way that can be very specific to a particular problem or optimization goal, such as analyzing the causes of memory bank conflicts or understanding an entire parallel algorithm.