What problem does it solve? Hand-written HIP kernels have no autotuner, so developers must choose launch geometry and compile flags themselves and often rely on flawed timing loops or guesswork about device limits. This Skill provides a disciplined workflow for selecting block/grid sizes against wave64, CU, and LDS constraints, measuring kernels correctly, and using rocprofv3 to verify which kernel actually ran. ## Core Features & Use Cases - Device-aware launch geometry: Read wavefront size, CU count, and LDS limits from rocminfo for gfx942 and gfx950 targets, and size blocks as multiples of 64 with grids covering all CUs. - Compile-flag search space: Use --offload-arch, launch_bounds, and -Rpass-analysis=kernel-resource-usage to prune register- and occupancy-doomed variants before benchmarking. - Ground-truth profiling with rocprofv3: Generate kernel stats CSVs that rank kernels by cost, confirm a tuned kernel engaged, and expose hidden costs like transposes and buffer fills. - Use Case: You ported a GEMM kernel from an NVIDIA GPU to MI300X and it underperforms. Use this Skill to correct the wave64 block sizing, synchronize and repeat your timing loop, then confirm via rocprofv3 that the intended kernel dominates runtime. ## Quick Start Ask the AI to tune your HIP kernel for your AMD Instinct GPU by reading device constants, fixing the launch geometry, and verifying the result with a rocprofv3 kernel trace.
