warp-specialization-report-skill

Profile a warp-specialized GPU kernel (CuTe-DSL or CUDA C++) by stamping the per-SM clock() register at each cross-warp synchronization point, then reconstruct a per-warp pipeline timeline (Gantt chart). Reveals where warps STALL, the cross-warp DATA DEPENDENCIES (which warp's output unblocks which), and how compute/memory phases OVERLAP. Use when the user wants to understand or visualize a warp-specialized kernel's runtime pipeline, find where warps wait on each other, or verify that producer/consumer warps actually run concurrently. Triggers include "profile the warp specialization", "pipeline timeline", "where does it stall", "how do the warps overlap", "visualize the producer/consumer pipeline". NOT applicable to Triton (cannot express warp specialization nor write timestamps to shared memory) and NOT a substitute for Nsight Compute/Systems for occupancy/throughput/memory-bandwidth analysis.