gpu-kernel-profile-optimizer

GPU kernel performance bottlenecks are identified using profiling tools like NCU and ROCprof to pinpoint exact issues such as memory throughput limits, low occupancy, or bank conflicts. This approach replaces intuition-based guessing with rigorous, evidence-based bottleneck analysis.

The best way to optimize a slow PyTorch operator is through a structured, profile-driven iterative workflow. This involves profiling the kernel, identifying specific warp stall reasons, and applying targeted optimizations like memory coalescing or register pressure reduction.

Yes, this profile-driven optimization workflow supports high-performance computing tasks on both NVIDIA and AMD hardware. It utilizes specific profiling tools like NCU for NVIDIA and ROCprof for AMD to measure performance and validate systematic code refinement.

To start an iterative GPU kernel optimization session, you provide the kernel file located in your current workspace to the skill. It then manages the entire lifecycle, including profiling, bottleneck analysis, plan generation, implementation, and validation.

An evidence-based optimization loop is necessary because it resolves hardware-level bottlenecks through rigorous validation and performance measurement. This structured cycle generates auditable evidence, ensuring systematic code refinement rather than relying on intuition-based guessing.

GPU kernel profiling can pinpoint exact hardware-level bottlenecks such as memory throughput limitations, insufficient occupancy, and bank conflicts. Identifying these specific issues allows for targeted code refinement during the optimization process.