kernel-profile

To diagnose if a CUDA kernel is memory-bound or compute-bound, you can profile it using Nsight Compute metrics to collect timing and stall data. This Skill validates kernel correctness first, then generates ncu_summary.md evidence to classify bottlenecks into memory, compute, latency, occupancy, or mixed behavior categories.

Yes, you can profile Triton kernels with Nsight Compute for evidence-backed tuning. This Skill supports Triton modules alongside CUDA-C++ and CUTLASS-wrapped kernels, collecting deterministic NCU metric summaries and detailed per-metric results to guide optimization directions like coalescing or pipeline utilization.

To run Nsight Compute profiling for bottleneck classification, you need the ncu and nsight-python readiness tooling installed. You also need a Python reference implementation defining reference(**kwargs) to validate correctness, and the numpy, torch, triton, and nvidia-cutlass-dsl dependencies available in your environment.

Validating kernel correctness before profiling works by running your kernel implementation against a Python reference function with configurable absolute and relative tolerance thresholds. The Skill executes this comparison and outputs a correctness.md report, ensuring your Nsight Compute metrics reflect an accurate kernel before you invest time in optimization.

Yes, this approach works with CUTLASS and CuTe DSL modules. The Skill confirms at least one usable backend—cuda-cpp, cute-dsl, cutlass, or triton—during environment checks, and applies its Nsight Compute profiling workflow to classify bottlenecks across these different GPU kernel implementations.

The best way to compare GPU kernel variants for occupancy and memory stalls is to profile them under fixed dimensions and settings using Nsight Compute. This Skill collects runtime timing via CUDA events and generates ncu_details.md, allowing you to evaluate different kernel implementations against consistent metrics.