ReProspect - A framework for reproducible prospecting of CUDA applications

Abstract

ReProspect is a Python framework for prospecting CUDA code, designed to ensure reproducibility through a fully programmatic approach.Prospecting encompasses three complementary ways of characterizing CUDA-based libraries and software components: how they interact with the CUDA runtime through API tracing, how kernels perform through kernel profiling, and how source constructs translate into machine code through binary analysis.ReProspect builds on NVIDIA tools: Nsight Systems, Nsight Compute, and the CUDA binary utilities.It streamlines data collection and extraction using these tools, and it complements them with new functionalities for programmatic analysis of these data, thus making it possible to encapsulate the entire prospecting analysis in a single Python script. Statement of needHPC software development strives to achieve performance and sustain it over time as hardware and software continue to evolve.Yet the modern programming landscape relies on complex software stacks and compiler toolchains that make it increasingly difficult to reason about code behavior.Developers therefore need tools that support continuous assessment of their code and help them unravel the implications of their design decisions and changes.Beyond ad hoc investigations, fully programmatic tools allow analyses to be integrated across the development cycle and open a range of new use cases.For instance, they support collaborative development by enabling developers to share concise, reproducible analyses that motivate design decisions and help reviewers grasp the impact of proposed changes.They also enable new types of tests in CI/CD pipelines that go beyond traditional output-correctness validation -such as confirming expected API call sequences for key library functionalities, verifying memory traffic for custom data layouts via kernel profiling, or validating the presence of specific instruction patterns in machine code.Finally, they can act as a framework for structuring research artifacts and documenting analyses, enabling others to reproduce and build upon prior work more effectively.For the CUDA stack, NVIDIA provides a set of proprietary tools guaranteed to be up-todate with their software and hardware.The runtime analysis tools Nsight Systems (NVIDIA Corporation, 2026d) and Nsight Compute (NVIDIA Corporation, 2026c) are designed for API tracing and kernel profiling, respectively.They both provide a GUI for exploring the results, as well as a low-level Python API for accessing the raw data.The CUDA binary utilities (NVIDIA Corporation, 2026a) provide command-line access to machine code (SASS or PTX (NVIDIA Corporation, 2026f)) and other information embedded in the binaries.However, while these tools allow raw data to be extracted, they do not themselves provide the infrastructure for effective programmatic analysis.