nvidia-cuda

nvidia-cuda

NVIDIA CUDA — 用于在 GPU 上执行通用并行计算的编程模型、工具链和库。 涵盖使用 CUDA C++/Python 编写内核、PTX ISA 指令集、面向 Blackwell 架构的调优， 以及使用 CUTLASS/CuTe 实现高性能 GEMM。

目录结构

skills/nvidia-cuda/
  SKILL.md
  references/
    programming-model/
      README.md
      introduction.md
      programming-model.md
      cuda-platform.md
      intro-to-cuda-cpp.md
      intro-to-cuda-python.md
      writing-cuda-kernels.md
      writing-tile-kernels.md
      asynchronous-execution.md
      understanding-memory.md
      nvcc.md
    advanced/
      README.md
      core/
        README.md
        advanced-host-programming.md
        advanced-kernel-programming.md
        driver-api.md
        multi-gpu-systems.md
        feature-survey.md
      features/
        README.md
        unified-memory.md
        cuda-graphs.md
        stream-ordered-memory-allocation.md
        cooperative-groups.md
        programmatic-dependent-launch.md
        green-contexts.md
        locality-domains.md
        compute-fabric-transport.md
        lazy-loading.md
        error-log-management.md
        async-barriers.md
        pipelines.md
        async-copies.md
        cluster-launch-control.md
        l2-cache-control.md
        memory-sync-domains.md
        inter-process-communication.md
        virtual-memory-management.md
        extended-gpu-memory.md
        dynamic-parallelism.md
        graphics-interop.md
        driver-entry-point-access.md
      appendices/
        README.md
        compute-capabilities.md
        environment-variables.md
        cpp-language-support.md
        cpp-language-extensions.md
        mathematical-functions.md
        device-callable-apis.md
        cuda-cpp-memory-model.md
        cuda-cpp-execution-model.md
    ptx-isa/
      README.md
      introduction.md
      programming-model.md
      machine-model.md
      syntax.md
      state-spaces-types-variables.md
      instruction-operands.md
      abi.md
      memory-consistency-model.md
      instruction-set-overview.md
      instructions-arithmetic.md
      instructions-comparison-logic.md
      instructions-data-movement.md
      instructions-texture-surface.md
      instructions-control-flow.md
      instructions-sync-communication.md
      instructions-matrix-multiply.md
      instructions-video-misc.md
      special-registers.md
      directives.md
      pragma-strings.md
      release-notes.md
    blackwell-tuning/
      README.md
      architecture.md
      best-practices.md
      streaming-multiprocessor.md
      memory-system.md
      nvlink.md
    cutlass/
      README.md
      cute-dsl/
        README.md
        overview.md
        functionality.md
        quick-start.md
        cute-dsl.md
        dsl-introduction.md
        dsl-code-generation.md
        dsl-control-flow.md
        dsl-jit-arg-generation.md
        dsl-dynamic-layout.md
        dsl-struct-types.md
        dsl-jit-caching.md
        dsl-jit-compilation-options.md
        dsl-types.md
        framework-integration.md
        debugging.md
        iket-profiling.md
        primitives-api.md
        task-scheduling.md
        ts-introduction.md
        ts-resources.md
        ts-pipelines.md
        ts-schedules.md
        ts-patterns.md
        ts-validation.md
        ts-printout.md
        ts-allocators.md
        ts-pipeline-groups.md
        ts-pdl.md
        ts-api-task-scheduling.md
        ts-api-resources.md
        ts-api-task.md
        ts-api-task-manager.md
        ts-api-schedule-builder.md
        ts-api-memory.md
        ts-api-pipeline-group.md
        ts-api-pipeline.md
        ts-api-enums.md
        autotuning-gemm.md
        compile-with-tvm-ffi.md
        dsl-ahead-of-time-compilation.md
        naming-conventions.md
        deprecation-policy.md
        mma-intro.md
        mma-wmma-programming.md
        mma-wgmma-programming.md
        mma-tcgen05-programming.md
        limitations.md
        faqs.md
        api-overview.md
        api-cute.md
        api-cute-arch.md
        api-cute-runtime.md
        api-cute-nvgpu.md
        api-cute-nvgpu-common.md
        api-cute-nvgpu-warp.md
        api-cute-nvgpu-warpgroup.md
        api-cute-nvgpu-cpasync.md
        api-cute-nvgpu-tcgen05.md
        api-pipeline.md
        api-utils.md
        api-utils-sm90.md
        api-utils-sm100.md
      operator-api/
        README.md
        overview.md
        tutorials.md
        tutorial-000-gemm.md
        tutorial-001-gemm-fused-epilogue.md
        tutorial-002-bring-your-own-kernel.md
        tutorial-003-host-latency-best-practices.md
        tutorial-004-fake-tensors.md
        tutorial-005-grouped-gemm-contiguous-offset.md
        tutorial-006-block-scaled-gemm.md
        tutorial-007-heuristics.md
        api-reference.md
        api-operator.md
        api-arguments.md
        api-discovery.md
        api-heuristics.md
        api-metadata.md
        api-misc.md
      cpp/
        README.md
        overview.md
        getting-started.md
        quickstart.md
        ide-setup.md
        build.md
        build-windows-visual-studio.md
        build-clang-host-compiler.md
        functionality.md
        terminology.md
        fundamental-types.md
        programming-guidelines.md
        gemm-heuristics.md
        efficient-gemm.md
        pipeline.md
        profiler.md
        gemm-performance-measurement.md
        dependent-kernel-launch.md
        blackwell.md
        blackwell-sm100-gemm.md
        blackwell-cluster-launch-control.md
        code-organization.md
        cute.md
        cute-quickstart.md
        cute-layout.md
        cute-layout-algebra.md
        cute-tensor.md
        cute-algorithms.md
        cute-mma-atom.md
        cute-gemm-tutorial.md
        cute-predication.md
        cute-tma-tensors.md
        cutlass-3x.md
        cutlass-3x-design.md
        cutlass-3x-backwards-compatibility.md
        gemm-api-3x.md
        cutlass-2x.md
        layout-2x.md
        gemm-api-2x.md
        tile-iterator-concept.md
        utilities.md
        grouped-scheduler.md
        implicit-gemm-convolution.md
  samples/
    README.md
    async-copy-pipeline.md
    cooperative-groups-reduction.md
    cuda-graph-explicit-nodes.md
    cuda-graph-stream-capture.md
    cute-dsl-gemm.md
    cutlass-cpp-device-gemm.md
    cutlass-operator-api-gemm.md
    explicit-memory-transfer.md
    multi-gpu-peer-to-peer.md
    shared-memory-tiled-matmul.md
    stream-ordered-allocation.md
    stream-overlap.md
    tile-kernel-vector-add.md
    unified-memory.md
    vector-add-kernel.md
  scripts/
    README.md
    install.md
    nvcc-build.md
    binary-utilities.md
    sanitize.md
    debug.md
    nsight-systems.md
    nsight-compute.md
    gpu-query.md
    cutlass-build.md

探索步骤

根据任务查找类别，并在类别的 README.md 中确定目标页面：

1. 在下方映射表中查找与任务对应的类别
2. 参阅该类别的 references/{category}/README.md（advanced/ cutlass/ 为子目录中的 README.md），确定目标页面
3. Read 相应的 .md 页面以了解详情

GB10 Grace Blackwell 实机的硬件规格、DGX OS、PXE/fleet 运维以及 playbook 启动步骤，请参阅 dgx-spark 技能； 远程连接管理、启动 VS Code 等桌面应用操作，请参阅 nvidia-sync 技能。 本技能负责 CUDA C++/Python 编写方式、PTX 指令和 CUTLASS API。

任务 → 类别映射

任务	类别	参考 README
想了解 CUDA 是什么、GPU 层级模型以及 CUDA 平台的整体概况	programming-model	references/programming-model/README.md
使用 CUDA C++/Python 编写内核、通过 nvcc 构建、处理异步执行和流	programming-model	references/programming-model/README.md
想了解 advanced 类别的整体结构（core / features / appendices）	advanced	references/advanced/README.md
想了解 Driver API、多 GPU、advanced host/kernel programming	advanced/core	references/advanced/core/README.md
想了解 CUDA Graphs / Cooperative Groups / Unified Memory / VMM / IPC / Green Contexts / Locality Domains / Compute Fabric Transport 等功能 API	advanced/features	references/advanced/features/README.md
想了解 Compute Capability、环境变量、C++ 语言扩展、内存/执行模型等技术附录	advanced/appendices	references/advanced/appendices/README.md
想了解 PTX 指令集、state space、寄存器、memory consistency model	ptx-isa	references/ptx-isa/README.md
想了解 Blackwell 架构、SM occupancy、NVLink 和内存系统调优	blackwell-tuning	references/blackwell-tuning/README.md
想了解 CUTLASS/CuTe 的整体类别结构（CuTe DSL / Operator API / C++）	cutlass	references/cutlass/README.md
想使用 CuTe DSL (Python) 编写 GEMM、MMA、布局和 Task Scheduling (TS)	cutlass/cute-dsl	references/cutlass/cute-dsl/README.md
想使用 CUTLASS Operator API（Python、调用已编译内核、通过 nvMatmulHeuristics 对 operator 排名）	cutlass/operator-api	references/cutlass/operator-api/README.md
想使用 CUTLASS C++ / CuTe C++ 实现基于模板的 GEMM 内核	cutlass/cpp	references/cutlass/cpp/README.md
想了解典型用法（vector add、tiled matmul、GEMM、CUDA Graphs 等）	samples	samples/README.md
想了解安装、nvcc 构建、nsight 性能分析以及 compute-sanitizer 命令	scripts	scripts/README.md