1. ROCm HIP 整体设计架构分析

本文ROCm hip库的结构设计解析。内容很多，可以当参考书来看。

1. 项目组织结构

层级	目录	说明
hip (公共仓库)	include/hip/	公共 API 头文件 (hip_runtime.h, hip_runtime_api.h, hiprtc.h)
	bin/	hipcc, hipconfig 工具
	cmake/	CMake 配置
CLR (实现仓库)	hipamd/src/	HIP AMD 后端实现 (hip_*.cpp, hip_internal.hpp)
	hipamd/hiprtc/	运行时编译实现
	rocclr/device/	设备抽象 (amd::Device 基类)
	rocclr/device/rocm/	ROCm 后端 (roc::Device → HSA API)
	rocclr/device/pal/	PAL 后端 (pal::Device)
	rocclr/platform/	平台抽象 (runtime, context, memory, command)
	rocclr/thread/	线程/同步原语
	opencl/	OpenCL 运行时 (共享 rocclr)

理解 HIP 时可以把它看成四个互相配合的面：

面向	主要内容	作用
API 兼容面	hip_runtime_api.h、hip_runtime.h	给应用提供 CUDA-like 的 C/C++ API
编译工具面	hipcc、CMake package、device headers	把 HIP 源码编译成 host object + GPU code object
运行时对象面	hip::Device、hip::Stream、hip::Event、memory/module/graph	在进程内维护 CUDA runtime 语义
后端执行面	ROCclr、ROCr、libhsakmt、KFD	把运行时对象转换成真实 GPU 队列、内存和命令

所以 projects/hip/ 更像"API 和工具的门面"，projects/clr/hipamd/ 才是 AMD 平台上 HIP Runtime 的主要实现，rocclr/ 则是 HIP 和 OpenCL 共用的设备运行时基础设施。

2. 分层架构

整个软件栈自上而下分为 API、抽象、后端、内核四段，用户的一次 API 调用会沿着这条链条层层下沉：
#mermaid-svg-lbt5AcGzTAo7so1L{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-lbt5AcGzTAo7so1L .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-lbt5AcGzTAo7so1L .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-lbt5AcGzTAo7so1L .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-lbt5AcGzTAo7so1L .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-lbt5AcGzTAo7so1L .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-lbt5AcGzTAo7so1L .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-lbt5AcGzTAo7so1L .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-lbt5AcGzTAo7so1L .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-lbt5AcGzTAo7so1L .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-lbt5AcGzTAo7so1L .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-lbt5AcGzTAo7so1L .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-lbt5AcGzTAo7so1L .marker.cross{stroke:#0b0b0b;}#mermaid-svg-lbt5AcGzTAo7so1L svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-lbt5AcGzTAo7so1L p{margin:0;}#mermaid-svg-lbt5AcGzTAo7so1L .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-lbt5AcGzTAo7so1L .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-lbt5AcGzTAo7so1L .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-lbt5AcGzTAo7so1L .cluster-label span p{background-color:transparent;}#mermaid-svg-lbt5AcGzTAo7so1L .label text,#mermaid-svg-lbt5AcGzTAo7so1L span{fill:#333;color:#333;}#mermaid-svg-lbt5AcGzTAo7so1L .node rect,#mermaid-svg-lbt5AcGzTAo7so1L .node circle,#mermaid-svg-lbt5AcGzTAo7so1L .node ellipse,#mermaid-svg-lbt5AcGzTAo7so1L .node polygon,#mermaid-svg-lbt5AcGzTAo7so1L .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-lbt5AcGzTAo7so1L .rough-node .label text,#mermaid-svg-lbt5AcGzTAo7so1L .node .label text,#mermaid-svg-lbt5AcGzTAo7so1L .image-shape .label,#mermaid-svg-lbt5AcGzTAo7so1L .icon-shape .label{text-anchor:middle;}#mermaid-svg-lbt5AcGzTAo7so1L .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-lbt5AcGzTAo7so1L .rough-node .label,#mermaid-svg-lbt5AcGzTAo7so1L .node .label,#mermaid-svg-lbt5AcGzTAo7so1L .image-shape .label,#mermaid-svg-lbt5AcGzTAo7so1L .icon-shape .label{text-align:center;}#mermaid-svg-lbt5AcGzTAo7so1L .node.clickable{cursor:pointer;}#mermaid-svg-lbt5AcGzTAo7so1L .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-lbt5AcGzTAo7so1L .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-lbt5AcGzTAo7so1L .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-lbt5AcGzTAo7so1L .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-lbt5AcGzTAo7so1L .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-lbt5AcGzTAo7so1L .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-lbt5AcGzTAo7so1L .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-lbt5AcGzTAo7so1L .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-lbt5AcGzTAo7so1L .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-lbt5AcGzTAo7so1L .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-lbt5AcGzTAo7so1L .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-lbt5AcGzTAo7so1L div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-lbt5AcGzTAo7so1L .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-lbt5AcGzTAo7so1L rect.text{fill:none;stroke-width:0;}#mermaid-svg-lbt5AcGzTAo7so1L .icon-shape,#mermaid-svg-lbt5AcGzTAo7so1L .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-lbt5AcGzTAo7so1L .icon-shape p,#mermaid-svg-lbt5AcGzTAo7so1L .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-lbt5AcGzTAo7so1L .icon-shape rect,#mermaid-svg-lbt5AcGzTAo7so1L .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-lbt5AcGzTAo7so1L .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-lbt5AcGzTAo7so1L .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-lbt5AcGzTAo7so1L :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} hipMalloc / hipMemcpy / hipMemAdvise
amd::Device 虚接口
后端实现
后端实现
用户应用

CUDA/HIP 代码
HIP API 层

hipamd/src/hip_*.cpp

参数校验 · 日志 · TLS 错误状态 · ihip* 内部实现
ROCclr 抽象层 rocclr/

amd::Device · amd::Memory

amd::Context · amd::CommandQueue

amd::Command 异步命令模型
rocm/ 后端

roc::Device → HSA API
pal/ 后端

pal::Device → PAL API
ROCr Runtime HSA

hsa_amd_svm_attributes_set 等
libhsakmt thunk

→ KFD ioctl
KFD 内核驱动

每一层只依赖它下面一层的抽象接口，因此后端（ROCm / PAL）可以整体替换而不影响上层 API。从设计上看，HIP 层不是直接把每个 API 翻译成 HSA 调用，而是在 hipamd 中先维护一套 CUDA 兼容的运行时对象模型，再把这些对象落到 ROCclr 的通用抽象上。三层对象的映射关系如下：
#mermaid-svg-CpdyThn16xCQ32OZ{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-CpdyThn16xCQ32OZ .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-CpdyThn16xCQ32OZ .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-CpdyThn16xCQ32OZ .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-CpdyThn16xCQ32OZ .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-CpdyThn16xCQ32OZ .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-CpdyThn16xCQ32OZ .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-CpdyThn16xCQ32OZ .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-CpdyThn16xCQ32OZ .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-CpdyThn16xCQ32OZ .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-CpdyThn16xCQ32OZ .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-CpdyThn16xCQ32OZ .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-CpdyThn16xCQ32OZ .marker.cross{stroke:#0b0b0b;}#mermaid-svg-CpdyThn16xCQ32OZ svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-CpdyThn16xCQ32OZ p{margin:0;}#mermaid-svg-CpdyThn16xCQ32OZ .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-CpdyThn16xCQ32OZ .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-CpdyThn16xCQ32OZ .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-CpdyThn16xCQ32OZ .cluster-label span p{background-color:transparent;}#mermaid-svg-CpdyThn16xCQ32OZ .label text,#mermaid-svg-CpdyThn16xCQ32OZ span{fill:#333;color:#333;}#mermaid-svg-CpdyThn16xCQ32OZ .node rect,#mermaid-svg-CpdyThn16xCQ32OZ .node circle,#mermaid-svg-CpdyThn16xCQ32OZ .node ellipse,#mermaid-svg-CpdyThn16xCQ32OZ .node polygon,#mermaid-svg-CpdyThn16xCQ32OZ .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-CpdyThn16xCQ32OZ .rough-node .label text,#mermaid-svg-CpdyThn16xCQ32OZ .node .label text,#mermaid-svg-CpdyThn16xCQ32OZ .image-shape .label,#mermaid-svg-CpdyThn16xCQ32OZ .icon-shape .label{text-anchor:middle;}#mermaid-svg-CpdyThn16xCQ32OZ .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-CpdyThn16xCQ32OZ .rough-node .label,#mermaid-svg-CpdyThn16xCQ32OZ .node .label,#mermaid-svg-CpdyThn16xCQ32OZ .image-shape .label,#mermaid-svg-CpdyThn16xCQ32OZ .icon-shape .label{text-align:center;}#mermaid-svg-CpdyThn16xCQ32OZ .node.clickable{cursor:pointer;}#mermaid-svg-CpdyThn16xCQ32OZ .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-CpdyThn16xCQ32OZ .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-CpdyThn16xCQ32OZ .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-CpdyThn16xCQ32OZ .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-CpdyThn16xCQ32OZ .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-CpdyThn16xCQ32OZ .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-CpdyThn16xCQ32OZ .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-CpdyThn16xCQ32OZ .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-CpdyThn16xCQ32OZ .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-CpdyThn16xCQ32OZ .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-CpdyThn16xCQ32OZ .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-CpdyThn16xCQ32OZ div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-CpdyThn16xCQ32OZ .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-CpdyThn16xCQ32OZ rect.text{fill:none;stroke-width:0;}#mermaid-svg-CpdyThn16xCQ32OZ .icon-shape,#mermaid-svg-CpdyThn16xCQ32OZ .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-CpdyThn16xCQ32OZ .icon-shape p,#mermaid-svg-CpdyThn16xCQ32OZ .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-CpdyThn16xCQ32OZ .icon-shape rect,#mermaid-svg-CpdyThn16xCQ32OZ .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-CpdyThn16xCQ32OZ .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-CpdyThn16xCQ32OZ .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-CpdyThn16xCQ32OZ :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 后端对象
ROCclr 对象
HIP Runtime 对象
hip::Device
hip::Stream
hip::Event
hipMalloc 指针
hipModule_t / kernel
hipGraph_t
amd::Context + amd::Device
amd::HostQueue
amd::Event / amd::Command
amd::Memory
amd::Program / device::Kernel
graph node DAG
roc::Device / pal::Device
roc::VirtualGPU / queue
HSA signal / HW event
VRAM/GTT/SVM allocation
code object / HSA executable
capture/replay 命令序列

因此 HIP 的职责可以概括为三件事：

1. 对外模拟 CUDA Runtime 语义：API 名称、错误码、默认 stream、device/context 行为尽量和 CUDA 对齐。
2. 对内维护 HIP 对象生命周期：device、stream、event、memory pool、module、graph 等对象都在 HIP 层先被组织起来。
3. 向下复用 ROCclr 能力：真正的设备、内存、队列、命令、事件由 ROCclr 承担，ROCm 后端再把这些抽象转换成 HSA/ROCr 操作。

3. HIP 层核心概念和关系

3.1 hip::Device: HIP 当前设备和 ROCclr Context 的桥

hip::Device 是 HIP 层最核心的运行时对象。它不是底层 GPU agent 本身，而是 HIP 对一个设备上下文的包装：

class Device : public amd::ReferenceCountedObject {
  amd::Context* context_;     // ROCclr context
  int deviceId_;              // HIP device ordinal
  Stream* null_stream_;       // legacy default stream
  MemoryPool* default_mem_pool_;
  ExecutionCtx* primaryExecCtx_;
};

关键关系：
#mermaid-svg-qKri2qfRRSVLf288{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-qKri2qfRRSVLf288 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-qKri2qfRRSVLf288 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-qKri2qfRRSVLf288 .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-qKri2qfRRSVLf288 .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-qKri2qfRRSVLf288 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-qKri2qfRRSVLf288 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-qKri2qfRRSVLf288 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-qKri2qfRRSVLf288 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-qKri2qfRRSVLf288 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-qKri2qfRRSVLf288 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-qKri2qfRRSVLf288 .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-qKri2qfRRSVLf288 .marker.cross{stroke:#0b0b0b;}#mermaid-svg-qKri2qfRRSVLf288 svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-qKri2qfRRSVLf288 p{margin:0;}#mermaid-svg-qKri2qfRRSVLf288 .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-qKri2qfRRSVLf288 .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-qKri2qfRRSVLf288 .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-qKri2qfRRSVLf288 .cluster-label span p{background-color:transparent;}#mermaid-svg-qKri2qfRRSVLf288 .label text,#mermaid-svg-qKri2qfRRSVLf288 span{fill:#333;color:#333;}#mermaid-svg-qKri2qfRRSVLf288 .node rect,#mermaid-svg-qKri2qfRRSVLf288 .node circle,#mermaid-svg-qKri2qfRRSVLf288 .node ellipse,#mermaid-svg-qKri2qfRRSVLf288 .node polygon,#mermaid-svg-qKri2qfRRSVLf288 .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-qKri2qfRRSVLf288 .rough-node .label text,#mermaid-svg-qKri2qfRRSVLf288 .node .label text,#mermaid-svg-qKri2qfRRSVLf288 .image-shape .label,#mermaid-svg-qKri2qfRRSVLf288 .icon-shape .label{text-anchor:middle;}#mermaid-svg-qKri2qfRRSVLf288 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-qKri2qfRRSVLf288 .rough-node .label,#mermaid-svg-qKri2qfRRSVLf288 .node .label,#mermaid-svg-qKri2qfRRSVLf288 .image-shape .label,#mermaid-svg-qKri2qfRRSVLf288 .icon-shape .label{text-align:center;}#mermaid-svg-qKri2qfRRSVLf288 .node.clickable{cursor:pointer;}#mermaid-svg-qKri2qfRRSVLf288 .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-qKri2qfRRSVLf288 .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-qKri2qfRRSVLf288 .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-qKri2qfRRSVLf288 .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-qKri2qfRRSVLf288 .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-qKri2qfRRSVLf288 .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-qKri2qfRRSVLf288 .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-qKri2qfRRSVLf288 .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-qKri2qfRRSVLf288 .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-qKri2qfRRSVLf288 .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-qKri2qfRRSVLf288 .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-qKri2qfRRSVLf288 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-qKri2qfRRSVLf288 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-qKri2qfRRSVLf288 rect.text{fill:none;stroke-width:0;}#mermaid-svg-qKri2qfRRSVLf288 .icon-shape,#mermaid-svg-qKri2qfRRSVLf288 .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-qKri2qfRRSVLf288 .icon-shape p,#mermaid-svg-qKri2qfRRSVLf288 .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-qKri2qfRRSVLf288 .icon-shape rect,#mermaid-svg-qKri2qfRRSVLf288 .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-qKri2qfRRSVLf288 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-qKri2qfRRSVLf288 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-qKri2qfRRSVLf288 :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} asContext
devices0
NullStream
memory pools
primaryExecCtx
hip::Device
amd::Context
amd::Device

硬件信息/内存属性/HSA agent
hip::Stream
hip::MemoryPool
hip::ExecutionCtx

这也解释了为什么 HIP API 经常先取：

hip::getCurrentDevice()->devices()[0]

前半段 hip::getCurrentDevice() 取的是 HIP 当前设备对象；后半段 devices()[0] 才进入 ROCclr 的 amd::Device，继续访问硬件信息、内存属性、HSA agent 等后端能力。

3.2 hip::Stream: HIP stream 对 ROCclr HostQueue 的包装

hip::Stream 继承自 amd::HostQueue，所以 HIP stream 本质上是 HIP 语义包装下的一条 ROCclr 命令队列：

class Stream : public amd::HostQueue {
  Device* device_;
  Priority priority_;
  unsigned int flags_;
  bool null_;
  uint64_t stream_id_;
};

它承载了几类语义：

HIP 语义	在 hip::Stream 中的体现
默认 stream	null_、NullStream()、PER_THREAD_DEFAULT_STREAM
stream priority	Priority::{High, Normal, Low}
non-blocking stream	flags_
stream capture	captureStatus_、pCaptureGraph_、captureID_
command ordering	继承 amd::HostQueue，向 ROCclr 队列提交命令

也就是说，HIP stream 是 CUDA 兼容语义和 ROCclr command queue 之间的适配层。

3.3 hip::Event: 同步点，底层绑定 ROCclr Event/Command

hip::Event 保存底层 amd::Event*，并通过 command/event 机制实现 record、wait、query、synchronize：

class Event {
  uint32_t flags_;
  amd::Event* event_;
  int device_id_;
};

典型关系是：
#mermaid-svg-T0TIUnbiMX9pz5Ez{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-T0TIUnbiMX9pz5Ez .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-T0TIUnbiMX9pz5Ez .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-T0TIUnbiMX9pz5Ez .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-T0TIUnbiMX9pz5Ez .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-T0TIUnbiMX9pz5Ez .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-T0TIUnbiMX9pz5Ez .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-T0TIUnbiMX9pz5Ez .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-T0TIUnbiMX9pz5Ez .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-T0TIUnbiMX9pz5Ez .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-T0TIUnbiMX9pz5Ez .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-T0TIUnbiMX9pz5Ez .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-T0TIUnbiMX9pz5Ez .marker.cross{stroke:#0b0b0b;}#mermaid-svg-T0TIUnbiMX9pz5Ez svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-T0TIUnbiMX9pz5Ez p{margin:0;}#mermaid-svg-T0TIUnbiMX9pz5Ez .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-T0TIUnbiMX9pz5Ez .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-T0TIUnbiMX9pz5Ez .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-T0TIUnbiMX9pz5Ez .cluster-label span p{background-color:transparent;}#mermaid-svg-T0TIUnbiMX9pz5Ez .label text,#mermaid-svg-T0TIUnbiMX9pz5Ez span{fill:#333;color:#333;}#mermaid-svg-T0TIUnbiMX9pz5Ez .node rect,#mermaid-svg-T0TIUnbiMX9pz5Ez .node circle,#mermaid-svg-T0TIUnbiMX9pz5Ez .node ellipse,#mermaid-svg-T0TIUnbiMX9pz5Ez .node polygon,#mermaid-svg-T0TIUnbiMX9pz5Ez .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-T0TIUnbiMX9pz5Ez .rough-node .label text,#mermaid-svg-T0TIUnbiMX9pz5Ez .node .label text,#mermaid-svg-T0TIUnbiMX9pz5Ez .image-shape .label,#mermaid-svg-T0TIUnbiMX9pz5Ez .icon-shape .label{text-anchor:middle;}#mermaid-svg-T0TIUnbiMX9pz5Ez .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-T0TIUnbiMX9pz5Ez .rough-node .label,#mermaid-svg-T0TIUnbiMX9pz5Ez .node .label,#mermaid-svg-T0TIUnbiMX9pz5Ez .image-shape .label,#mermaid-svg-T0TIUnbiMX9pz5Ez .icon-shape .label{text-align:center;}#mermaid-svg-T0TIUnbiMX9pz5Ez .node.clickable{cursor:pointer;}#mermaid-svg-T0TIUnbiMX9pz5Ez .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-T0TIUnbiMX9pz5Ez .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-T0TIUnbiMX9pz5Ez .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-T0TIUnbiMX9pz5Ez .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-T0TIUnbiMX9pz5Ez .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-T0TIUnbiMX9pz5Ez .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-T0TIUnbiMX9pz5Ez .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-T0TIUnbiMX9pz5Ez .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-T0TIUnbiMX9pz5Ez .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-T0TIUnbiMX9pz5Ez .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-T0TIUnbiMX9pz5Ez .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-T0TIUnbiMX9pz5Ez div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-T0TIUnbiMX9pz5Ez .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-T0TIUnbiMX9pz5Ez rect.text{fill:none;stroke-width:0;}#mermaid-svg-T0TIUnbiMX9pz5Ez .icon-shape,#mermaid-svg-T0TIUnbiMX9pz5Ez .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-T0TIUnbiMX9pz5Ez .icon-shape p,#mermaid-svg-T0TIUnbiMX9pz5Ez .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-T0TIUnbiMX9pz5Ez .icon-shape rect,#mermaid-svg-T0TIUnbiMX9pz5Ez .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-T0TIUnbiMX9pz5Ez .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-T0TIUnbiMX9pz5Ez .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-T0TIUnbiMX9pz5Ez :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} hipEventRecord

event, stream
在 hip::Stream 上

插入 marker/command
command 产生

amd::Event
hip::Event 绑定这个

amd::Event

所以 event 本身不是单独执行工作的实体，它是 stream 命令序列中的同步标记。

3.4 Memory: 指针 API 背后是 amd::Memory

HIP 对外暴露的是裸指针：

hipMalloc(&ptr, size);
hipMemcpy(dst, src, size, kind);
hipMemAdvise(ptr, size, advice, device);

但运行时内部需要把用户指针找回为 ROCclr 内存对象：

amd::Memory* getMemoryObject(hip::Device* device, const void* ptr, size_t& offset);
amd::MemObjMap::FindMemObj(ptr) -> amd::Memory*

这个设计很重要：HIP API 看起来是"指针式 API"，但 runtime 内部真正操作的是带元数据的 amd::Memory。元数据包括 allocation 类型、所属 context、设备位置、SVM/HMM 属性、offset 等。

3.5 Module / Kernel: code object 的加载和执行

HIP kernel 有两条来源：

1. 离线编译 ：hipcc 编译生成包含 code object 的可执行文件或 shared object
2. 运行时编译 ：libhiprtc.so 在运行时编译 HIP 源码，生成可加载 code object

无论来源如何，运行时执行都要经过 module/kernel 路径：
#mermaid-svg-92S9mYHaPppSWiKA{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-92S9mYHaPppSWiKA .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-92S9mYHaPppSWiKA .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-92S9mYHaPppSWiKA .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-92S9mYHaPppSWiKA .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-92S9mYHaPppSWiKA .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-92S9mYHaPppSWiKA .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-92S9mYHaPppSWiKA .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-92S9mYHaPppSWiKA .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-92S9mYHaPppSWiKA .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-92S9mYHaPppSWiKA .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-92S9mYHaPppSWiKA .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-92S9mYHaPppSWiKA .marker.cross{stroke:#0b0b0b;}#mermaid-svg-92S9mYHaPppSWiKA svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-92S9mYHaPppSWiKA p{margin:0;}#mermaid-svg-92S9mYHaPppSWiKA .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-92S9mYHaPppSWiKA .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-92S9mYHaPppSWiKA .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-92S9mYHaPppSWiKA .cluster-label span p{background-color:transparent;}#mermaid-svg-92S9mYHaPppSWiKA .label text,#mermaid-svg-92S9mYHaPppSWiKA span{fill:#333;color:#333;}#mermaid-svg-92S9mYHaPppSWiKA .node rect,#mermaid-svg-92S9mYHaPppSWiKA .node circle,#mermaid-svg-92S9mYHaPppSWiKA .node ellipse,#mermaid-svg-92S9mYHaPppSWiKA .node polygon,#mermaid-svg-92S9mYHaPppSWiKA .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-92S9mYHaPppSWiKA .rough-node .label text,#mermaid-svg-92S9mYHaPppSWiKA .node .label text,#mermaid-svg-92S9mYHaPppSWiKA .image-shape .label,#mermaid-svg-92S9mYHaPppSWiKA .icon-shape .label{text-anchor:middle;}#mermaid-svg-92S9mYHaPppSWiKA .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-92S9mYHaPppSWiKA .rough-node .label,#mermaid-svg-92S9mYHaPppSWiKA .node .label,#mermaid-svg-92S9mYHaPppSWiKA .image-shape .label,#mermaid-svg-92S9mYHaPppSWiKA .icon-shape .label{text-align:center;}#mermaid-svg-92S9mYHaPppSWiKA .node.clickable{cursor:pointer;}#mermaid-svg-92S9mYHaPppSWiKA .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-92S9mYHaPppSWiKA .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-92S9mYHaPppSWiKA .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-92S9mYHaPppSWiKA .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-92S9mYHaPppSWiKA .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-92S9mYHaPppSWiKA .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-92S9mYHaPppSWiKA .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-92S9mYHaPppSWiKA .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-92S9mYHaPppSWiKA .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-92S9mYHaPppSWiKA .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-92S9mYHaPppSWiKA .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-92S9mYHaPppSWiKA div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-92S9mYHaPppSWiKA .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-92S9mYHaPppSWiKA rect.text{fill:none;stroke-width:0;}#mermaid-svg-92S9mYHaPppSWiKA .icon-shape,#mermaid-svg-92S9mYHaPppSWiKA .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-92S9mYHaPppSWiKA .icon-shape p,#mermaid-svg-92S9mYHaPppSWiKA .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-92S9mYHaPppSWiKA .icon-shape rect,#mermaid-svg-92S9mYHaPppSWiKA .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-92S9mYHaPppSWiKA .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-92S9mYHaPppSWiKA .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-92S9mYHaPppSWiKA :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 执行
加载
符号供执行查找
hipModuleLoad / hipModuleLoadData
加载 code object
创建/查找 amd::Program
获取 kernel 符号
device::Kernel
hipModuleLaunchKernel / hipLaunchKernel
解析 grid/block/sharedMem/stream/args
创建 NDRangeKernelCommand
提交到 hip::Stream / amd::HostQueue

因此，HIP 层把 CUDA 风格的 kernel launch 参数转换成 ROCclr 能理解的 program/kernel/command 模型。

3.6 TLS: 当前设备、错误状态和默认 stream 的线程局部状态

HIP runtime 使用线程局部状态保存每个线程自己的 CUDA 兼容上下文：

class TlsAggregator {
  Device* device_;                   // 当前 device
  std::stack<Device*> ctxt_stack_;    // CUDA-style context stack
  hipError_t last_error_;             // sticky last error
  std::stack<ihipExec_t> exec_stack_;  // kernel launch 配置栈
  StreamPerThread stream_per_thread_obj_;
};

这解释了几个常见行为：

行为	HIP 内部依据
hipGetLastError() 每线程独立	thread_local hip::tls.last_error_
hipSetDevice() 只影响当前线程	thread_local hip::tls.device_
per-thread default stream	StreamPerThread 每线程懒创建
hipConfigureCall/hipSetupArgument 老式 launch API	exec_stack_ 保存临时 launch 配置

4. 典型 API 调用链

以几个常见 API 看 HIP 层如何把 CUDA 兼容接口落到 ROCclr。四条链虽然入口不同，但都遵循"参数校验 → 找对象 → 造命令/设属性 → 下沉后端"的统一套路：
#mermaid-svg-padqYdOlCtr5j0EI{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-padqYdOlCtr5j0EI .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-padqYdOlCtr5j0EI .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-padqYdOlCtr5j0EI .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-padqYdOlCtr5j0EI .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-padqYdOlCtr5j0EI .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-padqYdOlCtr5j0EI .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-padqYdOlCtr5j0EI .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-padqYdOlCtr5j0EI .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-padqYdOlCtr5j0EI .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-padqYdOlCtr5j0EI .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-padqYdOlCtr5j0EI .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-padqYdOlCtr5j0EI .marker.cross{stroke:#0b0b0b;}#mermaid-svg-padqYdOlCtr5j0EI svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-padqYdOlCtr5j0EI p{margin:0;}#mermaid-svg-padqYdOlCtr5j0EI .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-padqYdOlCtr5j0EI .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-padqYdOlCtr5j0EI .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-padqYdOlCtr5j0EI .cluster-label span p{background-color:transparent;}#mermaid-svg-padqYdOlCtr5j0EI .label text,#mermaid-svg-padqYdOlCtr5j0EI span{fill:#333;color:#333;}#mermaid-svg-padqYdOlCtr5j0EI .node rect,#mermaid-svg-padqYdOlCtr5j0EI .node circle,#mermaid-svg-padqYdOlCtr5j0EI .node ellipse,#mermaid-svg-padqYdOlCtr5j0EI .node polygon,#mermaid-svg-padqYdOlCtr5j0EI .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-padqYdOlCtr5j0EI .rough-node .label text,#mermaid-svg-padqYdOlCtr5j0EI .node .label text,#mermaid-svg-padqYdOlCtr5j0EI .image-shape .label,#mermaid-svg-padqYdOlCtr5j0EI .icon-shape .label{text-anchor:middle;}#mermaid-svg-padqYdOlCtr5j0EI .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-padqYdOlCtr5j0EI .rough-node .label,#mermaid-svg-padqYdOlCtr5j0EI .node .label,#mermaid-svg-padqYdOlCtr5j0EI .image-shape .label,#mermaid-svg-padqYdOlCtr5j0EI .icon-shape .label{text-align:center;}#mermaid-svg-padqYdOlCtr5j0EI .node.clickable{cursor:pointer;}#mermaid-svg-padqYdOlCtr5j0EI .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-padqYdOlCtr5j0EI .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-padqYdOlCtr5j0EI .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-padqYdOlCtr5j0EI .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-padqYdOlCtr5j0EI .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-padqYdOlCtr5j0EI .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-padqYdOlCtr5j0EI .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-padqYdOlCtr5j0EI .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-padqYdOlCtr5j0EI .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-padqYdOlCtr5j0EI .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-padqYdOlCtr5j0EI .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-padqYdOlCtr5j0EI div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-padqYdOlCtr5j0EI .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-padqYdOlCtr5j0EI rect.text{fill:none;stroke-width:0;}#mermaid-svg-padqYdOlCtr5j0EI .icon-shape,#mermaid-svg-padqYdOlCtr5j0EI .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-padqYdOlCtr5j0EI .icon-shape p,#mermaid-svg-padqYdOlCtr5j0EI .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-padqYdOlCtr5j0EI .icon-shape rect,#mermaid-svg-padqYdOlCtr5j0EI .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-padqYdOlCtr5j0EI .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-padqYdOlCtr5j0EI .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-padqYdOlCtr5j0EI :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 分配
拷贝
kernel
SVM 属性
hip* API 入口
HIP_INIT_API

参数校验 + 日志
ihip* 内部实现
hip::getCurrentDevice
getMemoryObject

找 ptr 对应 amd::Memory
动作类型
创建 amd::Memory

登记 ptr→Memory
Read/Write/CopyMemoryCommand
NDRangeKernelCommand
amd::Device::SetSvmAttributes
提交到 hip::Stream

= amd::HostQueue
后端转 HSA AQL packet
ROCm 后端 → HSA SVM 接口

下面逐个展开。

4.1 hipMalloc

#mermaid-svg-WzBJC828oFVsRcfM{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-WzBJC828oFVsRcfM .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-WzBJC828oFVsRcfM .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-WzBJC828oFVsRcfM .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-WzBJC828oFVsRcfM .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-WzBJC828oFVsRcfM .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-WzBJC828oFVsRcfM .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-WzBJC828oFVsRcfM .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-WzBJC828oFVsRcfM .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-WzBJC828oFVsRcfM .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-WzBJC828oFVsRcfM .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-WzBJC828oFVsRcfM .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-WzBJC828oFVsRcfM .marker.cross{stroke:#0b0b0b;}#mermaid-svg-WzBJC828oFVsRcfM svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-WzBJC828oFVsRcfM p{margin:0;}#mermaid-svg-WzBJC828oFVsRcfM .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-WzBJC828oFVsRcfM .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-WzBJC828oFVsRcfM .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-WzBJC828oFVsRcfM .cluster-label span p{background-color:transparent;}#mermaid-svg-WzBJC828oFVsRcfM .label text,#mermaid-svg-WzBJC828oFVsRcfM span{fill:#333;color:#333;}#mermaid-svg-WzBJC828oFVsRcfM .node rect,#mermaid-svg-WzBJC828oFVsRcfM .node circle,#mermaid-svg-WzBJC828oFVsRcfM .node ellipse,#mermaid-svg-WzBJC828oFVsRcfM .node polygon,#mermaid-svg-WzBJC828oFVsRcfM .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-WzBJC828oFVsRcfM .rough-node .label text,#mermaid-svg-WzBJC828oFVsRcfM .node .label text,#mermaid-svg-WzBJC828oFVsRcfM .image-shape .label,#mermaid-svg-WzBJC828oFVsRcfM .icon-shape .label{text-anchor:middle;}#mermaid-svg-WzBJC828oFVsRcfM .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-WzBJC828oFVsRcfM .rough-node .label,#mermaid-svg-WzBJC828oFVsRcfM .node .label,#mermaid-svg-WzBJC828oFVsRcfM .image-shape .label,#mermaid-svg-WzBJC828oFVsRcfM .icon-shape .label{text-align:center;}#mermaid-svg-WzBJC828oFVsRcfM .node.clickable{cursor:pointer;}#mermaid-svg-WzBJC828oFVsRcfM .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-WzBJC828oFVsRcfM .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-WzBJC828oFVsRcfM .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-WzBJC828oFVsRcfM .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-WzBJC828oFVsRcfM .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-WzBJC828oFVsRcfM .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-WzBJC828oFVsRcfM .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-WzBJC828oFVsRcfM .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-WzBJC828oFVsRcfM .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-WzBJC828oFVsRcfM .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-WzBJC828oFVsRcfM .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-WzBJC828oFVsRcfM div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-WzBJC828oFVsRcfM .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-WzBJC828oFVsRcfM rect.text{fill:none;stroke-width:0;}#mermaid-svg-WzBJC828oFVsRcfM .icon-shape,#mermaid-svg-WzBJC828oFVsRcfM .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-WzBJC828oFVsRcfM .icon-shape p,#mermaid-svg-WzBJC828oFVsRcfM .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-WzBJC828oFVsRcfM .icon-shape rect,#mermaid-svg-WzBJC828oFVsRcfM .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-WzBJC828oFVsRcfM .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-WzBJC828oFVsRcfM .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-WzBJC828oFVsRcfM :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} hipMalloc

ptr, size
HIP_INIT_API

参数检查
ihipMalloc

ptr, size, flags
hip::getCurrentDevice
amd::Context/Device

创建 amd::Memory
返回裸指针

内部登记 ptr→amd::Memory

4.2 hipMemcpyAsync

#mermaid-svg-Vt3u2LICW0qn8Gbv{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-Vt3u2LICW0qn8Gbv .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-Vt3u2LICW0qn8Gbv .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-Vt3u2LICW0qn8Gbv .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-Vt3u2LICW0qn8Gbv .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-Vt3u2LICW0qn8Gbv .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-Vt3u2LICW0qn8Gbv .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-Vt3u2LICW0qn8Gbv .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-Vt3u2LICW0qn8Gbv .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-Vt3u2LICW0qn8Gbv .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-Vt3u2LICW0qn8Gbv .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-Vt3u2LICW0qn8Gbv .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-Vt3u2LICW0qn8Gbv .marker.cross{stroke:#0b0b0b;}#mermaid-svg-Vt3u2LICW0qn8Gbv svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-Vt3u2LICW0qn8Gbv p{margin:0;}#mermaid-svg-Vt3u2LICW0qn8Gbv .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-Vt3u2LICW0qn8Gbv .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-Vt3u2LICW0qn8Gbv .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-Vt3u2LICW0qn8Gbv .cluster-label span p{background-color:transparent;}#mermaid-svg-Vt3u2LICW0qn8Gbv .label text,#mermaid-svg-Vt3u2LICW0qn8Gbv span{fill:#333;color:#333;}#mermaid-svg-Vt3u2LICW0qn8Gbv .node rect,#mermaid-svg-Vt3u2LICW0qn8Gbv .node circle,#mermaid-svg-Vt3u2LICW0qn8Gbv .node ellipse,#mermaid-svg-Vt3u2LICW0qn8Gbv .node polygon,#mermaid-svg-Vt3u2LICW0qn8Gbv .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-Vt3u2LICW0qn8Gbv .rough-node .label text,#mermaid-svg-Vt3u2LICW0qn8Gbv .node .label text,#mermaid-svg-Vt3u2LICW0qn8Gbv .image-shape .label,#mermaid-svg-Vt3u2LICW0qn8Gbv .icon-shape .label{text-anchor:middle;}#mermaid-svg-Vt3u2LICW0qn8Gbv .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-Vt3u2LICW0qn8Gbv .rough-node .label,#mermaid-svg-Vt3u2LICW0qn8Gbv .node .label,#mermaid-svg-Vt3u2LICW0qn8Gbv .image-shape .label,#mermaid-svg-Vt3u2LICW0qn8Gbv .icon-shape .label{text-align:center;}#mermaid-svg-Vt3u2LICW0qn8Gbv .node.clickable{cursor:pointer;}#mermaid-svg-Vt3u2LICW0qn8Gbv .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-Vt3u2LICW0qn8Gbv .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-Vt3u2LICW0qn8Gbv .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-Vt3u2LICW0qn8Gbv .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-Vt3u2LICW0qn8Gbv .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-Vt3u2LICW0qn8Gbv .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-Vt3u2LICW0qn8Gbv .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-Vt3u2LICW0qn8Gbv .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-Vt3u2LICW0qn8Gbv .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-Vt3u2LICW0qn8Gbv .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-Vt3u2LICW0qn8Gbv .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-Vt3u2LICW0qn8Gbv div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-Vt3u2LICW0qn8Gbv .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-Vt3u2LICW0qn8Gbv rect.text{fill:none;stroke-width:0;}#mermaid-svg-Vt3u2LICW0qn8Gbv .icon-shape,#mermaid-svg-Vt3u2LICW0qn8Gbv .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-Vt3u2LICW0qn8Gbv .icon-shape p,#mermaid-svg-Vt3u2LICW0qn8Gbv .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-Vt3u2LICW0qn8Gbv .icon-shape rect,#mermaid-svg-Vt3u2LICW0qn8Gbv .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-Vt3u2LICW0qn8Gbv .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-Vt3u2LICW0qn8Gbv .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-Vt3u2LICW0qn8Gbv :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} hipMemcpyAsync

dst, src, size, kind, stream
解析 hipStream_t

空 stream→default stream
getMemoryObject

找 src/dst 的 amd::Memory
创建 Read/Write/CopyMemoryCommand
提交到 hip::Stream

= amd::HostQueue

4.3 hipLaunchKernel

#mermaid-svg-VmBzCE0ZTCCjVukv{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-VmBzCE0ZTCCjVukv .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-VmBzCE0ZTCCjVukv .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-VmBzCE0ZTCCjVukv .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-VmBzCE0ZTCCjVukv .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-VmBzCE0ZTCCjVukv .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-VmBzCE0ZTCCjVukv .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-VmBzCE0ZTCCjVukv .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-VmBzCE0ZTCCjVukv .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-VmBzCE0ZTCCjVukv .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-VmBzCE0ZTCCjVukv .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-VmBzCE0ZTCCjVukv .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-VmBzCE0ZTCCjVukv .marker.cross{stroke:#0b0b0b;}#mermaid-svg-VmBzCE0ZTCCjVukv svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-VmBzCE0ZTCCjVukv p{margin:0;}#mermaid-svg-VmBzCE0ZTCCjVukv .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-VmBzCE0ZTCCjVukv .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-VmBzCE0ZTCCjVukv .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-VmBzCE0ZTCCjVukv .cluster-label span p{background-color:transparent;}#mermaid-svg-VmBzCE0ZTCCjVukv .label text,#mermaid-svg-VmBzCE0ZTCCjVukv span{fill:#333;color:#333;}#mermaid-svg-VmBzCE0ZTCCjVukv .node rect,#mermaid-svg-VmBzCE0ZTCCjVukv .node circle,#mermaid-svg-VmBzCE0ZTCCjVukv .node ellipse,#mermaid-svg-VmBzCE0ZTCCjVukv .node polygon,#mermaid-svg-VmBzCE0ZTCCjVukv .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-VmBzCE0ZTCCjVukv .rough-node .label text,#mermaid-svg-VmBzCE0ZTCCjVukv .node .label text,#mermaid-svg-VmBzCE0ZTCCjVukv .image-shape .label,#mermaid-svg-VmBzCE0ZTCCjVukv .icon-shape .label{text-anchor:middle;}#mermaid-svg-VmBzCE0ZTCCjVukv .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-VmBzCE0ZTCCjVukv .rough-node .label,#mermaid-svg-VmBzCE0ZTCCjVukv .node .label,#mermaid-svg-VmBzCE0ZTCCjVukv .image-shape .label,#mermaid-svg-VmBzCE0ZTCCjVukv .icon-shape .label{text-align:center;}#mermaid-svg-VmBzCE0ZTCCjVukv .node.clickable{cursor:pointer;}#mermaid-svg-VmBzCE0ZTCCjVukv .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-VmBzCE0ZTCCjVukv .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-VmBzCE0ZTCCjVukv .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-VmBzCE0ZTCCjVukv .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-VmBzCE0ZTCCjVukv .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-VmBzCE0ZTCCjVukv .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-VmBzCE0ZTCCjVukv .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-VmBzCE0ZTCCjVukv .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-VmBzCE0ZTCCjVukv .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-VmBzCE0ZTCCjVukv .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-VmBzCE0ZTCCjVukv .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-VmBzCE0ZTCCjVukv div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-VmBzCE0ZTCCjVukv .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-VmBzCE0ZTCCjVukv rect.text{fill:none;stroke-width:0;}#mermaid-svg-VmBzCE0ZTCCjVukv .icon-shape,#mermaid-svg-VmBzCE0ZTCCjVukv .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-VmBzCE0ZTCCjVukv .icon-shape p,#mermaid-svg-VmBzCE0ZTCCjVukv .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-VmBzCE0ZTCCjVukv .icon-shape rect,#mermaid-svg-VmBzCE0ZTCCjVukv .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-VmBzCE0ZTCCjVukv .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-VmBzCE0ZTCCjVukv .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-VmBzCE0ZTCCjVukv :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} hipLaunchKernel

kernel, grid, block, args, sharedMem, stream
解析当前 device 和 stream
查找 kernel 对应的

device::Kernel
打包 kernel arguments
创建 NDRangeKernelCommand
stream 提交

后端转 HSA AQL packet

4.4 hipMemAdvise

#mermaid-svg-UfhpoPqmDFDsBO7d{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-UfhpoPqmDFDsBO7d .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-UfhpoPqmDFDsBO7d .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-UfhpoPqmDFDsBO7d .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-UfhpoPqmDFDsBO7d .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-UfhpoPqmDFDsBO7d .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-UfhpoPqmDFDsBO7d .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-UfhpoPqmDFDsBO7d .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-UfhpoPqmDFDsBO7d .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-UfhpoPqmDFDsBO7d .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-UfhpoPqmDFDsBO7d .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-UfhpoPqmDFDsBO7d .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-UfhpoPqmDFDsBO7d .marker.cross{stroke:#0b0b0b;}#mermaid-svg-UfhpoPqmDFDsBO7d svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-UfhpoPqmDFDsBO7d p{margin:0;}#mermaid-svg-UfhpoPqmDFDsBO7d .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-UfhpoPqmDFDsBO7d .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-UfhpoPqmDFDsBO7d .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-UfhpoPqmDFDsBO7d .cluster-label span p{background-color:transparent;}#mermaid-svg-UfhpoPqmDFDsBO7d .label text,#mermaid-svg-UfhpoPqmDFDsBO7d span{fill:#333;color:#333;}#mermaid-svg-UfhpoPqmDFDsBO7d .node rect,#mermaid-svg-UfhpoPqmDFDsBO7d .node circle,#mermaid-svg-UfhpoPqmDFDsBO7d .node ellipse,#mermaid-svg-UfhpoPqmDFDsBO7d .node polygon,#mermaid-svg-UfhpoPqmDFDsBO7d .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-UfhpoPqmDFDsBO7d .rough-node .label text,#mermaid-svg-UfhpoPqmDFDsBO7d .node .label text,#mermaid-svg-UfhpoPqmDFDsBO7d .image-shape .label,#mermaid-svg-UfhpoPqmDFDsBO7d .icon-shape .label{text-anchor:middle;}#mermaid-svg-UfhpoPqmDFDsBO7d .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-UfhpoPqmDFDsBO7d .rough-node .label,#mermaid-svg-UfhpoPqmDFDsBO7d .node .label,#mermaid-svg-UfhpoPqmDFDsBO7d .image-shape .label,#mermaid-svg-UfhpoPqmDFDsBO7d .icon-shape .label{text-align:center;}#mermaid-svg-UfhpoPqmDFDsBO7d .node.clickable{cursor:pointer;}#mermaid-svg-UfhpoPqmDFDsBO7d .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-UfhpoPqmDFDsBO7d .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-UfhpoPqmDFDsBO7d .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-UfhpoPqmDFDsBO7d .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-UfhpoPqmDFDsBO7d .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-UfhpoPqmDFDsBO7d .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-UfhpoPqmDFDsBO7d .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-UfhpoPqmDFDsBO7d .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-UfhpoPqmDFDsBO7d .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-UfhpoPqmDFDsBO7d .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-UfhpoPqmDFDsBO7d .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-UfhpoPqmDFDsBO7d div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-UfhpoPqmDFDsBO7d .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-UfhpoPqmDFDsBO7d rect.text{fill:none;stroke-width:0;}#mermaid-svg-UfhpoPqmDFDsBO7d .icon-shape,#mermaid-svg-UfhpoPqmDFDsBO7d .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-UfhpoPqmDFDsBO7d .icon-shape p,#mermaid-svg-UfhpoPqmDFDsBO7d .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-UfhpoPqmDFDsBO7d .icon-shape rect,#mermaid-svg-UfhpoPqmDFDsBO7d .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-UfhpoPqmDFDsBO7d .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-UfhpoPqmDFDsBO7d .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-UfhpoPqmDFDsBO7d :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} hipMemAdvise

ptr, count, advice, device
HIP_INIT_API
ihipMemAdvise
getMemoryObject ptr

找 amd::Memory
选择目标 amd::Device
amd::Device::SetSvmAttributes
ROCm 后端调用

HSA SVM 属性接口

这条路径体现了 HIP SVM/HMM 设计的关键点：API 是按指针范围传入的，但真正的状态最终落到进程级 SVM range / ROCr / KFD 管理上。

5. 核心设计模式

5.1 设备抽象 (amd::Device)

// rocclr/device/device.hpp
class Device : public RuntimeObject {
public:
  // 后端实现的纯虚接口
  virtual bool SetSvmAttributes(...) const = 0;
  virtual Memory* createBuffer(...) = 0;
  virtual CommandQueue* createQueue(...) = 0;
  // ...
};

// rocclr/device/rocm/rocdevice.hpp
class roc::Device : public amd::Device {
  // ROCm/HSA 后端实现
  hsa_agent_t agent_;  // HSA agent handle
};

5.2 命令模型 (amd::Command)

ROCclr 采用异步命令队列模型，所有设备操作都被建模为 amd::Command 的派生类，入队后延迟执行：
#mermaid-svg-0ayEr2Xt7KPuktGd{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-0ayEr2Xt7KPuktGd .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-0ayEr2Xt7KPuktGd .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-0ayEr2Xt7KPuktGd .error-icon{fill:#552222;}#mermaid-svg-0ayEr2Xt7KPuktGd .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-0ayEr2Xt7KPuktGd .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-0ayEr2Xt7KPuktGd .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-0ayEr2Xt7KPuktGd .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-0ayEr2Xt7KPuktGd .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-0ayEr2Xt7KPuktGd .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-0ayEr2Xt7KPuktGd .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-0ayEr2Xt7KPuktGd .marker{fill:#333333;stroke:#333333;}#mermaid-svg-0ayEr2Xt7KPuktGd .marker.cross{stroke:#333333;}#mermaid-svg-0ayEr2Xt7KPuktGd svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-0ayEr2Xt7KPuktGd p{margin:0;}#mermaid-svg-0ayEr2Xt7KPuktGd .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-0ayEr2Xt7KPuktGd .cluster-label text{fill:#333;}#mermaid-svg-0ayEr2Xt7KPuktGd .cluster-label span{color:#333;}#mermaid-svg-0ayEr2Xt7KPuktGd .cluster-label span p{background-color:transparent;}#mermaid-svg-0ayEr2Xt7KPuktGd .label text,#mermaid-svg-0ayEr2Xt7KPuktGd span{fill:#333;color:#333;}#mermaid-svg-0ayEr2Xt7KPuktGd .node rect,#mermaid-svg-0ayEr2Xt7KPuktGd .node circle,#mermaid-svg-0ayEr2Xt7KPuktGd .node ellipse,#mermaid-svg-0ayEr2Xt7KPuktGd .node polygon,#mermaid-svg-0ayEr2Xt7KPuktGd .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-0ayEr2Xt7KPuktGd .rough-node .label text,#mermaid-svg-0ayEr2Xt7KPuktGd .node .label text,#mermaid-svg-0ayEr2Xt7KPuktGd .image-shape .label,#mermaid-svg-0ayEr2Xt7KPuktGd .icon-shape .label{text-anchor:middle;}#mermaid-svg-0ayEr2Xt7KPuktGd .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-0ayEr2Xt7KPuktGd .rough-node .label,#mermaid-svg-0ayEr2Xt7KPuktGd .node .label,#mermaid-svg-0ayEr2Xt7KPuktGd .image-shape .label,#mermaid-svg-0ayEr2Xt7KPuktGd .icon-shape .label{text-align:center;}#mermaid-svg-0ayEr2Xt7KPuktGd .node.clickable{cursor:pointer;}#mermaid-svg-0ayEr2Xt7KPuktGd .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-0ayEr2Xt7KPuktGd .arrowheadPath{fill:#333333;}#mermaid-svg-0ayEr2Xt7KPuktGd .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-0ayEr2Xt7KPuktGd .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-0ayEr2Xt7KPuktGd .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-0ayEr2Xt7KPuktGd .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-0ayEr2Xt7KPuktGd .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-0ayEr2Xt7KPuktGd .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-0ayEr2Xt7KPuktGd .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-0ayEr2Xt7KPuktGd .cluster text{fill:#333;}#mermaid-svg-0ayEr2Xt7KPuktGd .cluster span{color:#333;}#mermaid-svg-0ayEr2Xt7KPuktGd div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-0ayEr2Xt7KPuktGd .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-0ayEr2Xt7KPuktGd rect.text{fill:none;stroke-width:0;}#mermaid-svg-0ayEr2Xt7KPuktGd .icon-shape,#mermaid-svg-0ayEr2Xt7KPuktGd .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-0ayEr2Xt7KPuktGd .icon-shape p,#mermaid-svg-0ayEr2Xt7KPuktGd .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-0ayEr2Xt7KPuktGd .icon-shape rect,#mermaid-svg-0ayEr2Xt7KPuktGd .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-0ayEr2Xt7KPuktGd .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-0ayEr2Xt7KPuktGd .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-0ayEr2Xt7KPuktGd :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} amd::Command

异步命令基类
ReadMemoryCommand
WriteMemoryCommand
CopyMemoryCommand
NDRangeKernelCommand
SvmPrefetchAsyncCommand

SVM 相关
... 其他命令

5.3 内存对象跟踪

// 全局虚拟地址 → amd::Memory 映射
amd::MemObjMap::FindMemObj(ptr)  → amd::Memory*

6. 关键设计决策

设计决策	目的
头文件分离	hip/ 只有公共 API，实现在 clr/
ROCclr 抽象	同一套代码支持 HIP + OpenCL
后端插件化	rocm/ vs pal/ 可切换
动态符号加载	ROCclr 不硬链接 ROCr，运行时 dlopen
TLS 错误状态	hip::tls.last_error_ 线程安全
命令队列模型	异步执行，延迟提交
HIP 对象包装 ROCclr 对象	保持 CUDA 语义，同时复用 ROCclr 的 device/context/memory/queue
裸指针 API + 内部内存表	对外保持 CUDA 风格，对内用 amd::Memory 保存 allocation 元数据
当前 device 是线程局部状态	hipSetDevice、默认 stream、last error 都遵循 CUDA 的 per-thread 行为
RTC 独立成库	普通 runtime 程序不必加载运行时编译逻辑，和 CUDA nvrtc 模型一致
Graph/Capture 建在 stream 之上	capture 不是单独的后端通路，而是记录 stream 上的命令依赖关系

7. 构建产物

libamdhip64.so          ← HIP 运行时库
libhiprtc.so            ← 运行时编译库
libhiprtc-builtins.so   ← 内置函数

在当前 CLR Debug 构建中，这些库会生成在：

projects/clr/build/hipamd/lib/

实际生成结果通常是一组带版本号的真实文件和无版本软链接：

libamdhip64.so
  -> libamdhip64.so.7
  -> libamdhip64.so.7.14.60850-<git-hash>

libhiprtc.so
  -> libhiprtc.so.7
  -> libhiprtc.so.7.14.60850-<git-hash>

libhiprtc-builtins.so
  -> libhiprtc-builtins.so.7
  -> libhiprtc-builtins.so.7.14.60850-<git-hash>

7.1 libamdhip64.so

libamdhip64.so 是 HIP Runtime 的主库，对应 CUDA 里的 libcudart.so。普通 HIP 程序调用的 hipMalloc、hipMemcpy、hipLaunchKernel、hipMemAdvise 等运行时 API，最终主要进入这个库。

它的源码主体在：

projects/clr/hipamd/src/

例如：

hip_memory.cpp   # 内存相关 API
hip_device.cpp   # 设备相关 API
hip_stream.cpp   # stream/queue 相关 API
hip_module.cpp   # module/kernel 加载相关 API

7.2 libhiprtc.so

libhiprtc.so 是 HIP Runtime Compilation 库，对应 CUDA 里的 libnvrtc.so。它负责在程序运行时把 HIP kernel 源码编译成可加载的代码对象。

典型 API 包括：

hiprtcCreateProgram(...);
hiprtcCompileProgram(...);
hiprtcGetCode(...);
hiprtcDestroyProgram(...);

也就是说，普通预编译 HIP 程序不一定直接依赖 libhiprtc.so；只有使用运行时编译能力的程序才需要显式链接或加载它。

源码主体在：

projects/clr/hipamd/src/hiprtc/

HIPRTC 曾经和 HIP Runtime 符号混在一起，后来被拆成独立库。这样设计的好处是：

1. 普通 HIP Runtime 程序不必加载运行时编译相关逻辑
2. 和 CUDA 的 cudart / nvrtc 分离模型一致
3. hiprtc 可以独立暴露运行时编译 API 和 CMake package

7.3 libhiprtc-builtins.so

libhiprtc-builtins.so 是 HIPRTC 编译时使用的内置函数库，服务于运行时编译流程。它不是普通 HIP Runtime API 的主入口，而是 libhiprtc.so 在编译 kernel 源码时需要配合使用的组件。

8. 与 CUDA 的对应关系

CUDA	HIP	说明
libcudart.so	libamdhip64.so	运行时库
nvcc	hipcc	编译器驱动
cudaMemAdvise	hipMemAdvise	API 1:1 映射
CUcontext	hip::Device	上下文/设备

更完整的概念对应关系如下：

CUDA 概念	HIP 表达	HIP 内部对象	ROCclr/ROCr 落点
Runtime API	hip* API	hipamd/src/hip_*.cpp	ROCclr command/device/memory
Device ordinal	int device	hip::Device::deviceId_	amd::Device / HSA agent
Context	primary context / current device	hip::Device + TLS context stack	amd::Context
Stream	hipStream_t	hip::Stream	amd::HostQueue
Event	hipEvent_t	hip::Event	amd::Event / HSA signal
Device memory	hipMalloc pointer	amd::Memory	VRAM/GTT allocation
Managed memory	hipMallocManaged pointer	amd::Memory + SVM/HMM metadata	ROCr SVM / KFD SVM range
Module	hipModule_t	loaded code object/program	amd::Program
Kernel function	hipFunction_t / host stub	kernel metadata	device::Kernel
Graph	hipGraph_t	HIP graph DAG	captured command sequence
RTC	hiprtc* API	HIPRTC program object	clang/comgr/code object

9. HIP API 实现模式

9.1 API 入口宏

// hipamd/src/hip_internal.hpp
#define HIP_INIT_API(cid, ...)                           \
  HIP_INIT_CB(cid, args);                                \
  amd::Thread* thread = amd::Thread::current();          \
  if (!VDI_CHECK_THREAD(thread)) {                       \
    HIP_RETURN(hipErrorOutOfMemory);                     \
  }                                                      \
  HIP_CB_SPAWNER_OBJECT(cid);

9.2 内部实现分离

HIP 普遍采用"公共 API 外壳 + ihip* 内部实现"的分离模式：外壳负责初始化/校验/错误码包装，内部函数才写真正的业务逻辑：
#mermaid-svg-xhMof13GVZmWOgxt{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-xhMof13GVZmWOgxt .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-xhMof13GVZmWOgxt .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-xhMof13GVZmWOgxt .error-icon{fill:hsl(220.5882352941, 100%, 98.3333333333%);}#mermaid-svg-xhMof13GVZmWOgxt .error-text{fill:rgb(8.5000000002, 5.7500000001, 0);stroke:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-xhMof13GVZmWOgxt .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-xhMof13GVZmWOgxt .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-xhMof13GVZmWOgxt .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-xhMof13GVZmWOgxt .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-xhMof13GVZmWOgxt .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-xhMof13GVZmWOgxt .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-xhMof13GVZmWOgxt .marker{fill:#0b0b0b;stroke:#0b0b0b;}#mermaid-svg-xhMof13GVZmWOgxt .marker.cross{stroke:#0b0b0b;}#mermaid-svg-xhMof13GVZmWOgxt svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:22px;}#mermaid-svg-xhMof13GVZmWOgxt p{margin:0;}#mermaid-svg-xhMof13GVZmWOgxt .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-xhMof13GVZmWOgxt .cluster-label text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-xhMof13GVZmWOgxt .cluster-label span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-xhMof13GVZmWOgxt .cluster-label span p{background-color:transparent;}#mermaid-svg-xhMof13GVZmWOgxt .label text,#mermaid-svg-xhMof13GVZmWOgxt span{fill:#333;color:#333;}#mermaid-svg-xhMof13GVZmWOgxt .node rect,#mermaid-svg-xhMof13GVZmWOgxt .node circle,#mermaid-svg-xhMof13GVZmWOgxt .node ellipse,#mermaid-svg-xhMof13GVZmWOgxt .node polygon,#mermaid-svg-xhMof13GVZmWOgxt .node path{fill:#fff4dd;stroke:hsl(40.5882352941, 60%, 83.3333333333%);stroke-width:1px;}#mermaid-svg-xhMof13GVZmWOgxt .rough-node .label text,#mermaid-svg-xhMof13GVZmWOgxt .node .label text,#mermaid-svg-xhMof13GVZmWOgxt .image-shape .label,#mermaid-svg-xhMof13GVZmWOgxt .icon-shape .label{text-anchor:middle;}#mermaid-svg-xhMof13GVZmWOgxt .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-xhMof13GVZmWOgxt .rough-node .label,#mermaid-svg-xhMof13GVZmWOgxt .node .label,#mermaid-svg-xhMof13GVZmWOgxt .image-shape .label,#mermaid-svg-xhMof13GVZmWOgxt .icon-shape .label{text-align:center;}#mermaid-svg-xhMof13GVZmWOgxt .node.clickable{cursor:pointer;}#mermaid-svg-xhMof13GVZmWOgxt .root .anchor path{fill:#0b0b0b!important;stroke-width:0;stroke:#0b0b0b;}#mermaid-svg-xhMof13GVZmWOgxt .arrowheadPath{fill:#0b0b0b;}#mermaid-svg-xhMof13GVZmWOgxt .edgePath .path{stroke:#0b0b0b;stroke-width:2.0px;}#mermaid-svg-xhMof13GVZmWOgxt .flowchart-link{stroke:#0b0b0b;fill:none;}#mermaid-svg-xhMof13GVZmWOgxt .edgeLabel{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-xhMof13GVZmWOgxt .edgeLabel p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-xhMof13GVZmWOgxt .edgeLabel rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-xhMof13GVZmWOgxt .labelBkg{background-color:rgba(243.9999999999, 220.9999999998, 255, 0.5);}#mermaid-svg-xhMof13GVZmWOgxt .cluster rect{fill:hsl(220.5882352941, 100%, 98.3333333333%);stroke:hsl(220.5882352941, 60%, 88.3333333333%);stroke-width:1px;}#mermaid-svg-xhMof13GVZmWOgxt .cluster text{fill:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-xhMof13GVZmWOgxt .cluster span{color:rgb(8.5000000002, 5.7500000001, 0);}#mermaid-svg-xhMof13GVZmWOgxt div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(220.5882352941, 100%, 98.3333333333%);border:1px solid hsl(220.5882352941, 60%, 88.3333333333%);border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-xhMof13GVZmWOgxt .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-xhMof13GVZmWOgxt rect.text{fill:none;stroke-width:0;}#mermaid-svg-xhMof13GVZmWOgxt .icon-shape,#mermaid-svg-xhMof13GVZmWOgxt .image-shape{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);text-align:center;}#mermaid-svg-xhMof13GVZmWOgxt .icon-shape p,#mermaid-svg-xhMof13GVZmWOgxt .image-shape p{background-color:hsl(-79.4117647059, 100%, 93.3333333333%);padding:2px;}#mermaid-svg-xhMof13GVZmWOgxt .icon-shape rect,#mermaid-svg-xhMof13GVZmWOgxt .image-shape rect{opacity:0.5;background-color:hsl(-79.4117647059, 100%, 93.3333333333%);fill:hsl(-79.4117647059, 100%, 93.3333333333%);}#mermaid-svg-xhMof13GVZmWOgxt .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-xhMof13GVZmWOgxt .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-xhMof13GVZmWOgxt :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} 用户调用

hipMemAdvise
公共 API 外壳

HIP_INIT_API + HIP_RETURN
ihipMemAdvise

内部实现/业务逻辑
ROCclr / 后端

// 公共 API
hipError_t hipMemAdvise(const void* ptr, size_t count, 
                        hipMemoryAdvise advice, int device) {
  HIP_INIT_API(hipMemAdvise, ptr, count, advice, device);
  HIP_RETURN(ihipMemAdvise(ptr, count, advice, device));  // 调用内部实现
}

// 内部实现
hipError_t ihipMemAdvise(const void* ptr, size_t count,
                         hipMemoryAdvise advice, int device) {
  // 实际业务逻辑
}

9.3 版本后缀 (_v2)

// hip_runtime_api.h
#define hipMemAdvise hipMemAdvise_v2

// 实现
hipError_t hipMemAdvise_v2(...);  // 新版本
hipError_t hipMemAdvise_v1(...);  // 旧版本 (兼容)

10. 源码目录详解

10.1 projects/hip/ (公共头文件)

include/hip/
├── hip_runtime.h          # 主入口，包含所有 HIP 功能
├── hip_runtime_api.h      # 运行时 API 声明
├── hip_vector_types.h     # float4, int2 等向量类型
├── hip_fp16.h             # 半精度浮点支持
├── hip_cooperative_groups.h  # 协作组
├── hiprtc.h               # 运行时编译 API
└── driver_types.h         # CUDA 驱动兼容类型

10.2 projects/clr/hipamd/ (HIP AMD 实现)

src/
├── hip_memory.cpp         # 内存管理 API
├── hip_device.cpp         # 设备管理 API
├── hip_stream.cpp         # 流/队列 API
├── hip_event.cpp          # 事件 API
├── hip_module.cpp         # 模块/内核加载
├── hip_texture.cpp        # 纹理 API
├── hip_graph*.cpp         # CUDA Graph 兼容
└── hip_internal.hpp       # 内部宏和工具

10.3 projects/clr/rocclr/ (ROCclr 抽象层)

device/
├── device.hpp             # amd::Device 基类
├── rocm/                  # ROCm/HSA 后端
│   ├── rocdevice.hpp      # roc::Device
│   ├── rocmemory.hpp      # roc::Memory
│   └── rocvirtual.hpp     # roc::VirtualGPU
└── pal/                   # PAL 后端 (Windows/专业卡)

platform/
├── runtime.hpp            # 运行时初始化
├── context.hpp            # amd::Context
├── memory.hpp             # amd::Memory
├── command.hpp            # amd::Command
└── commandqueue.hpp       # amd::CommandQueue

11. 总结

把前面所有调用链收敛成一条端到端主干，一次用户 API 从进入到落到内核的完整旅程如下：
#mermaid-svg-xAs0HcuW4ImOvnw4{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-xAs0HcuW4ImOvnw4 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-xAs0HcuW4ImOvnw4 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-xAs0HcuW4ImOvnw4 .error-icon{fill:#552222;}#mermaid-svg-xAs0HcuW4ImOvnw4 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-xAs0HcuW4ImOvnw4 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-xAs0HcuW4ImOvnw4 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-xAs0HcuW4ImOvnw4 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-xAs0HcuW4ImOvnw4 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-xAs0HcuW4ImOvnw4 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-xAs0HcuW4ImOvnw4 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-xAs0HcuW4ImOvnw4 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-xAs0HcuW4ImOvnw4 .marker.cross{stroke:#333333;}#mermaid-svg-xAs0HcuW4ImOvnw4 svg{font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-xAs0HcuW4ImOvnw4 p{margin:0;}#mermaid-svg-xAs0HcuW4ImOvnw4 .label{font-family:"trebuchet ms",verdana,arial,sans-serif;color:#333;}#mermaid-svg-xAs0HcuW4ImOvnw4 .cluster-label text{fill:#333;}#mermaid-svg-xAs0HcuW4ImOvnw4 .cluster-label span{color:#333;}#mermaid-svg-xAs0HcuW4ImOvnw4 .cluster-label span p{background-color:transparent;}#mermaid-svg-xAs0HcuW4ImOvnw4 .label text,#mermaid-svg-xAs0HcuW4ImOvnw4 span{fill:#333;color:#333;}#mermaid-svg-xAs0HcuW4ImOvnw4 .node rect,#mermaid-svg-xAs0HcuW4ImOvnw4 .node circle,#mermaid-svg-xAs0HcuW4ImOvnw4 .node ellipse,#mermaid-svg-xAs0HcuW4ImOvnw4 .node polygon,#mermaid-svg-xAs0HcuW4ImOvnw4 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-xAs0HcuW4ImOvnw4 .rough-node .label text,#mermaid-svg-xAs0HcuW4ImOvnw4 .node .label text,#mermaid-svg-xAs0HcuW4ImOvnw4 .image-shape .label,#mermaid-svg-xAs0HcuW4ImOvnw4 .icon-shape .label{text-anchor:middle;}#mermaid-svg-xAs0HcuW4ImOvnw4 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-xAs0HcuW4ImOvnw4 .rough-node .label,#mermaid-svg-xAs0HcuW4ImOvnw4 .node .label,#mermaid-svg-xAs0HcuW4ImOvnw4 .image-shape .label,#mermaid-svg-xAs0HcuW4ImOvnw4 .icon-shape .label{text-align:center;}#mermaid-svg-xAs0HcuW4ImOvnw4 .node.clickable{cursor:pointer;}#mermaid-svg-xAs0HcuW4ImOvnw4 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-xAs0HcuW4ImOvnw4 .arrowheadPath{fill:#333333;}#mermaid-svg-xAs0HcuW4ImOvnw4 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-xAs0HcuW4ImOvnw4 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-xAs0HcuW4ImOvnw4 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-xAs0HcuW4ImOvnw4 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-xAs0HcuW4ImOvnw4 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-xAs0HcuW4ImOvnw4 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-xAs0HcuW4ImOvnw4 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-xAs0HcuW4ImOvnw4 .cluster text{fill:#333;}#mermaid-svg-xAs0HcuW4ImOvnw4 .cluster span{color:#333;}#mermaid-svg-xAs0HcuW4ImOvnw4 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:"trebuchet ms",verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-xAs0HcuW4ImOvnw4 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-xAs0HcuW4ImOvnw4 rect.text{fill:none;stroke-width:0;}#mermaid-svg-xAs0HcuW4ImOvnw4 .icon-shape,#mermaid-svg-xAs0HcuW4ImOvnw4 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-xAs0HcuW4ImOvnw4 .icon-shape p,#mermaid-svg-xAs0HcuW4ImOvnw4 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-xAs0HcuW4ImOvnw4 .icon-shape rect,#mermaid-svg-xAs0HcuW4ImOvnw4 .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-xAs0HcuW4ImOvnw4 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-xAs0HcuW4ImOvnw4 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-xAs0HcuW4ImOvnw4 :root{--mermaid-font-family:"trebuchet ms",verdana,arial,sans-serif;} rocm/
pal/
用户 hip* API
TLS: 当前 device/默认 stream/错误状态
HIP 对象层

hip::Device/Stream/Event/Memory
ROCclr 抽象

amd::Context/HostQueue/Command/Memory
后端
roc::Device → HSA
pal::Device → PAL
ROCr Runtime
libhsakmt → KFD ioctl
GPU 硬件队列/内存

HIP 的设计遵循以下核心原则：

1. CUDA 源码兼容：API 命名和语义尽量与 CUDA 一致
2. 分层解耦：HIP API → ROCclr → 后端 → HSA → KFD
3. 后端可插拔：同一 ROCclr 支持多种后端 (ROCm, PAL)
4. 运行时共享：HIP 和 OpenCL 共享 ROCclr 层
5. 延迟加载：HSA 符号运行时 dlopen，减少启动开销