Visualization of profiling and tracing in CPU‐GPU programs

Abstract

Summary As the complexity of the toolchain increases for heterogeneous CPU‐GPU systems, the needs for comprehensive tracing and debugging tools also grows. Heterogeneous platforms bring new possibilities but also new performance issues that are hard to detect. Some techniques that were used on CPU programs are now adapted to GPUs. However, there are some concepts specific to GPUs, like SIMD processing, and the effects of the close interactions between the CPUs and the GPUs, with shared virtual memory and user‐level queues. Multiple sources of data need to be extracted and correlated to obtain a more global view of the performance. In this article, we introduce a novel approach for measuring and visualizing performance defects inside CPU‐GPU programs by combining kernel events, compute kernel events, user API calls and memory transfers. We created two new views that combine this information, to help provide a global view. This framework uses the open source user queue system described in the HSA standard. It can easily be adapted to any user queue system for heterogeneous computing devices. We compare this framework with current existing tools and test it against the Rodinia benchmark. We look at how the execution behavior affects the tracing and profiling overhead and we use Trace Compass to visualize the resulting trace.