cuda-cpp-kernel

CUDA C++ and PTX Kernel Development

Goal

Use the bundled CUDA, PTX, and profiling references in this skill to implement, debug, and optimize CUDA kernels without relying on agent-specific install paths or ad hoc web searches.

When to Use

Use this skill when you need to:

• write or review CUDA C++ kernels or supporting host code
• reason about PTX instructions, inline PTX, Tensor Core instructions, or memory model details
• debug CUDA Runtime API or Driver API failures
• profile or optimize a kernel with Nsight Systems, Nsight Compute, compute-sanitizer, or cuda-gdb
• investigate shared memory bank conflicts, memory coalescing, occupancy, register pressure, async copy, TMA, or cluster behavior
• compare a custom operator or kernel against a PyTorch baseline for correctness or performance

Do not use this skill for:

• CUTLASS or CuTe template design as the primary task; use cutlass-cpp-kernel
• CuTe DSL Python kernel authoring as the primary task; use cute-dsl-kernel
• cache-dit operator registration, packaging, or public API migration work by itself; pair with operator-migration

Reference Style Rule

Use skill-local relative paths for bundled references, for example:

• references/ptx-docs/
• references/cuda-runtime-docs/
• references/ncu-docs/ProfilingGuide.md

Do not write agent-specific install paths into follow-up notes or generated docs.

Bundled Reference Map

The following directories are bundled under references/ inside this skill:

• references/ptx-docs/ — full PTX ISA reference
• references/ptx-simple/ — condensed PTX quick reference
• references/cuda-runtime-docs/ — CUDA Runtime API reference
• references/cuda-driver-docs/ — CUDA Driver API reference
• references/cuda-guide/ — CUDA Programming Guide
• references/best-practices-guide/ — CUDA C++ Best Practices Guide
• references/ncu-docs/ — Nsight Compute docs
• references/nsys-docs/ — Nsight Systems docs
• references/debugging-tools.md — debugging workflow notes
• references/performance-traps.md — common optimization traps