kernel-profile

What problem does it solve?

kernel-profile helps you determine whether a CUDA/CUTLASS/CuTe DSL/Triton kernel is memory-bound, compute-bound, latency-bound, occupancy-limited, or mixed by validating correctness first and then collecting Nsight Compute (NCU) metrics with nsight-python.

Core Features & Use Cases

• Environment readiness & backend availability checks: verifies CUDA/PyTorch/GPU selection plus readiness of ncu and nsight-python, and confirms at least one backend is usable (cuda-cpp, cute-dsl, cutlass, or triton).
• Correctness validation before profiling: runs the provided kernel implementation against a Python reference (with configurable atol/rtol) and writes correctness.md.
• Deterministic NCU profiling workflow: collects kernel runtime timing (CUDA events) plus NCU metric summaries and detailed per-metric results, then classifies bottlenecks using the generated ncu_summary.md evidence.
• Use cases: diagnosing slow GPU kernels, choosing the next optimization direction (e.g., coalescing/misalignment vs compute pipeline utilization vs synchronization vs divergence), and comparing occupancy/memory stalls across kernel variants under fixed dimensions and settings.

Quick Start

Use the kernel-profile skill to validate correctness and generate NCU profiling reports for your kernel implementation against a Python reference, then read ncu_summary.md to decide the bottleneck category.