domain-skills

Choose the next CUDA optimization experiment from diagnosed bottlenecks. Use when profiling has already identified the hotspot and, when needed, attributed it to a source region, and the task is to inspect that CUDA source and decide which fix family or next experiment is most justified. This includes optimization strategies that deliberately trade speed against numerical fidelity through a lower-precision path once the format contract is understood, and standard-library-style CUDA primitive substitution when a local hotspot reimplements a known primitive. Do not use this skill for hotspot finding, profiler-to-source association, running profilers, generic CUDA code review, big-picture kernel or hot-path redesign, or workload-domain pattern selection; route those to `krnopt-cuda-profiling`, `krnopt-cuda-coding`, `krnopt-cuda-structural-optimization`, or `krnopt-cuda-domain-optimization`.

Domain-specific CUDA kernel optimization guidance for choosing and planning kernel designs by workload family. Use when Codex needs to optimize a CUDA kernel using domain patterns such as MoE routing, grouped GEMM, fused expert MLPs, sparse dispatch/combine, persistent MoE scheduling, padding elimination, or future domain-specific CUDA playbooks under this skill.

Reason about architecture-specific CUDA optimization. Use when a user wants to exploit hardware features on a target GPU, map features to code structure, enhance GPU utilization through hardware-specific surfaces, compare hardware-specific optimization directions, or produce a hardware-aware optimization plan. Route generic cross-evidence optimization directions to `krnopt-cuda-generic-optimization`, big-picture kernel or hot-path redesign to `krnopt-cuda-structural-optimization`, workload-domain pattern selection to `krnopt-cuda-domain-optimization`, generic CUDA coding doctrine to `krnopt-cuda-coding`, profiler-first diagnosis to `krnopt-cuda-profiling`, and low-precision format-contract questions to `krnopt-low-precision-kernel-formats`.

Write good, correct, efficient CUDA source code. Use when the task is authoring or modifying `.cu` code or CUDA device helpers and the main need is to choose a sound ownership model, make correctness explicit, shape memory access and reuse well, manage shared memory and register tradeoffs, keep synchronization and control flow sane, and choose dtypes or instruction paths deliberately. Do not use this skill for profiler-driven diagnosis, benchmark interpretation, ranked optimization planning, or big-picture kernel or hot-path redesign; route measured-evidence analysis to `krnopt-cuda-profiling`, next-experiment planning to `krnopt-cuda-generic-optimization`, and structural replanning to `krnopt-cuda-structural-optimization`.

Identify CUDA bottlenecks from source code and profiling evidence. Use when the task is to inspect CUDA source, analyze `nsys` or `ncu` profiling data, determine which kernel or code region is actually limiting performance, and attribute the bottleneck back to specific source blocks, loops, memory paths, or source lines when the evidence supports it. Do not use this skill for choosing optimization directions or rewriting the code; route localized source-plus-evidence optimization planning to `krnopt-cuda-generic-optimization`, broader redesign handoffs to `krnopt-cuda-structural-optimization`, and diagnosed MoE or other covered domain-pattern choices to `krnopt-cuda-domain-optimization`.

Replan and redesign CUDA kernels from a big-picture view. Use when the user wants a structural change to a kernel or hot path, such as rethinking boundaries, scheduling, primitives, pipeline stages, or metadata flow, instead of pursuing another local optimization. Do targeted idea search or source research before guessing when deeper knowledge is needed, and route workload-domain pattern selection such as MoE routing, grouped GEMM, dispatch/combine, expert MLP fusion, padding elimination, or persistent MoE scheduling to `krnopt-cuda-domain-optimization`. Do not use this skill for hotspot finding, raw profiler interpretation, or low-level CUDA coding.

Use and understand low-precision CUDA kernel formats. Use when the task is to understand, compare, choose, or apply FP8, block-scaled FP8, NVFP4, MXFP4/6/8, dynamic quantization, or related low-precision kernel formats; interpret the scale-layout, packing, dtype-path, or dequant contract a kernel must obey; or decide which other CUDA skill should handle profiling, optimization direction, structural redesign, hardware-specific format choice, or source implementation next.