With the widespread application of graphics processing units (GPUs) in high-performance computing, artificial intelligence, and other fields, the importance of GPU performance analysis has become increasingly prominent. However, issues arising from GPUs’ parallel processing model, complex scheduling mechanisms, multi-level memory systems, and heterogeneous CPU-GPU communication make such performance analysis extremely difficult. This study first reviews the GPU execution model and the roofline model for GPU performance analysis and summarizes typical GPU performance bottlenecks. Based on a review of existing research paradigms, a classification system for GPU performance analysis methods is proposed, with implementation mechanisms as its guiding principle and analysis granularity and tool forms as its objectives. At the macro level, these methods are divided into five categories, including methods based on hardware performance counters, methods based on micro-benchmarks, methods based on software instrumentation, methods based on simulation and modeling, and methods based on binary instrumentation. On this basis, this study discusses the advantages and shortcomings of each type of method and technique. In particular, with Neutrino as an example, the analysis reveals the limitations of traditional tools in observing microscopic scheduling behavior. Finally, the current challenges faced by GPU performance analysis are summarized, and the development directions of next-generation performance analysis tools, including programmability, fine granularity, and intelligence, are identified.