In Situ Code Profiling IDE Feedback for CUDA Applications

The evolving landscape of software development has shifted from optimizing for hardware limitations to prioritizing functionality, maintainability, and developer productivity. While resource abundance has allowed a more balanced approach, certain applications necessitate highly performant code. Software profiling, particularly in optimization, addresses bottlenecks, enhancing program efficiency. The spatial disconnect between profiling results and code impact challenges developers, motivating integration solutions like Performance-Hat[10] for CPUs. The absence of a GPU-focused tool prompts the exploration of a similar approach for Nvidia GPUs, using the Fine-Grained Memory Profiler[7] as the foundational profiling tool for integration into the Integrated Development Environment (IDE). In this thesis, a profiling tool is proposed that is effective for examining memory transactions, with a particular focus on caching behavior and visualizing the cache hit rate. Compatible with various Nvidia GPUs for profiling CUDA code, it provides an easy-to-use experience without requiring complicated knowledge of their inner workings. Visualization techniques that allow users to identify behavioral anomalies in source code.