ROCm Performance Guard

ROCm Performance Guard is an evidence-driven performance analysis toolkit that helps developers identify performance regressions, verify numerical correctness, and evaluate optimization candidates using reproducible measurements and predefined acceptance rules. The project combines Python-based benchmarking and analysis modules with controlled GPU experiments. It collects environment information, compares benchmark configurations, investigates performance changes, recommends optimization experiments, and evaluates results using repeated measurements, variability analysis, correctness checks, and preserved experimental evidence. The demonstrated AMD GPU optimization experiments use native Windows OpenCL on an integrated AMD Radeon GPU (gfx90c), not ROCm/HIP. In GEMM experiments, 8×8 local-memory tiling reduced measured median kernel latency from 13.71730 ms to 6.13649 ms. Adding column-first work-item mapping reduced it further to 3.71454 ms, achieving 72.2661 GFLOPS. Additional loop-unrolling candidates were rejected because they failed to meet the predefined improvement threshold. Selected tests were also run in AMD Developer Cloud, where the environment reported an AMD Instinct MI300X virtual function and ROCm-related tools. A total of 169 selected tests passed, validating selected analysis, testing, and capability-detection paths. A separate standalone HIP vector-add test passed, but it was not the project's GEMM implementation. The native ROCm/HIP GEMM backend remains unimplemented. The next goal is to implement and validate it on supported ROCm hardware while maintaining transparent measurements, correctness checks, and reproducible optimization criteria.