amd-rocm

amd-rocm

AMD ROCm——用于在 Instinct/Radeon GPU 上进行通用并行计算的软件栈。 涵盖通过 HIP 编程模型和运行时 API 编写内核、rocBLAS/MIOpen/RCCL 等库， 以及面向 CDNA/RDNA 架构的调优。

适用范围：ROCm Core SDK 10.1.0（TheRock，2026-10-05）/ HIP 7.16.0（从 baseline 7.14.0 更新）。

目录结构

skills/amd-rocm/
  SKILL.md
  references/
    rocm-stack/
      README.md
      what-is-rocm.md
      core-sdk-overview.md
      math-compute-libraries.md
      communication-libraries.md
      runtime-and-compilers.md
      profiling-debugging-tools.md
      control-monitoring-tools.md
      media-storage-libraries.md
      install-linux.md
      install-deep-learning-frameworks.md
      install-docker-spack.md
      install-windows.md
      compatibility-and-release-history.md
      ai-ecosystem.md
      gpu-systems-infra.md
      toolkits.md
      threading-libraries.md
      libhipcxx.md
    hip/
      README.md
      programming-model.md
      hardware-implementation.md
      compilers-clr.md
      performance-guidelines.md
      cpp-language-extensions.md
      kernel-language-cpp-support.md
      hiprtc.md
      debugging-logging.md
      runtime-api-overview.md
      initialization-version.md
      device-management.md
      error-handling.md
      stream-management.md
      event-management.md
      asynchronous-execution.md
      execution-control-launch.md
      occupancy.md
      module-context-management.md
      memory-management.md
      host-memory.md
      device-memory.md
      unified-memory.md
      virtual-memory.md
      stream-ordered-allocator.md
      texture-surface.md
      graph-management.md
      cooperative-groups.md
      multi-device-peer-to-peer.md
      interop.md
      porting-cuda-to-hip.md
      cuda-to-hip-api-comparison.md
      hipify.md
      math-api.md
      env-variables.md
      deprecated-apis.md
      hardware-features.md
      execution-context.md
      complex-math-api.md
      low-precision-types.md
      device-properties-cdna.md
    gpu-arch/
      README.md
      gpu-architecture-overview.md
      gpu-arch-specs.md
      cdna-architecture.md
      rdna-architecture.md
      performance-optimization-guides.md
      mi100-microarchitecture.md
      mi250-microarchitecture.md
      mi300-microarchitecture.md
      precision-support.md
      mi300-mi200-performance-counters.md
      mi350-performance-counters.md
      isa-documentation.md
      machine-readable-isa.md
  samples/
    README.md
    vector-add-kernel.md
    explicit-memory-transfer.md
    shared-memory-tiled-matmul.md
    dynamic-shared-memory.md
    unified-memory.md
    stream-ordered-allocation.md
    stream-overlap.md
    hip-graph-stream-capture.md
    hip-graph-explicit-nodes.md
    cooperative-groups-reduction.md
    warp-shuffle-reduction.md
    multi-gpu-peer-to-peer.md
    occupancy-tuning.md
    hipify-cuda-to-hip.md
    event-timing.md
    device-query.md
    hiprtc-runtime-compilation.md
    rocblas-gemm.md
  scripts/
    README.md
    install.md
    hipcc-build.md
    hipify.md
    gpu-query.md
    profiling.md
    debug.md
    sanitize.md
    frameworks.md
    validation.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（samples/ 和 scripts/ 则参阅其根目录下的 README.md），确定目标页面
3. Read 对应的 .md 文件，查看详情

CUDA C++/Python 内核编写、PTX ISA 指令集以及使用 CUTLASS/CuTe 实现 GEMM，请参阅 nvidia-cuda 技能。 本技能负责 HIP 编程模型、运行时 API、ROCm 库以及 CDNA/RDNA 架构。 由于 HIP 的命名与 CUDA API 几乎一一对应（例如，hipMalloc ↔ cudaMalloc）， references/hip/cuda-to-hip-api-comparison.md 和 references/hip/porting-cuda-to-hip.md 汇总了 HIP 与 CUDA 的对应关系及迁移步骤。

任务 → 类别映射

任务	类别	参考 README
想了解 ROCm 软件栈全貌、Core SDK 与 Extras 的区别以及 TheRock 构建系统	rocm-stack	references/rocm-stack/README.md
想了解 rocBLAS/MIOpen/RCCL 等数值计算和通信库、hipThreads（GPU threading）、libhipcxx、媒体/视觉（RPP）以及分析/监控工具（rocprofv3、amd-smi）	rocm-stack	references/rocm-stack/README.md
想了解 Linux/Windows 安装步骤（TheRock amdrocm 软件包、pip / tarball / runfile）、PyTorch/JAX 安装（TensorFlow 2.19.1 已删除）以及 GPU/OS 兼容性	rocm-stack	references/rocm-stack/README.md
想了解 HIP 编程模型、内核编写、C++ 语言扩展以及 HIPRTC 运行时编译	hip	references/hip/README.md
想了解 HIP 运行时 API（设备/流/事件/内存管理、HIP Graph、Cooperative Groups）	hip	references/hip/README.md
想了解 HIP 执行上下文（hipGreenCtxCreate、CU 分区）、复数数学 API、低精度 fp8/fp6/fp4 类型以及 CDNA 设备拓扑（XCD/XCC）	hip	references/hip/README.md
想了解从 CUDA 移植到 HIP、使用 hipify 转换代码以及 CUDA↔HIP API 对照表	hip	references/hip/README.md
想了解 CDNA/RDNA 架构设计、MI100/MI250/MI300 微架构以及 ISA/性能计数器	gpu-arch	references/gpu-arch/README.md
想了解典型用法（vector add、HIP Graph、rocBLAS GEMM、hipify 转换等）	samples	samples/README.md
想了解安装、hipcc 构建、rocminfo/amd-smi、rocprofv3 分析以及 rocgdb 调试命令	scripts	scripts/README.md