展示如何使用msdebug工具来上板调试一个PyTorch接口调用的add算子，该add算子可实现两个向量相加并输出结果的功能。

前提条件

已安装配套的CANN软件包并配置环境变量，完成开发环境和运行环境的准备。
单击Link获取样例工程，为进行算子上板调优、仿真调优及调试做准备。
已参考《Ascend Extension for PyTorch 配置与安装》，完成PyTorch框架和torch_npu插件的安装。

操作步骤

编译算子。

进入AddCustom算子工程，把AddCustom/CMakePresets.json中的编译配置从Release修改为Debug。

# CMakePresets.json
...
    "configurePresets": [
        {
            "name": "default",
            ...
                "CMAKE_BUILD_TYPE": {
                    "type": "STRING",
                    "value": "Debug" # 编译配置修改为Debug
                }

执行/build.sh命令完成算子的编译和部署，编译成功后，会在当前目录下创建build_out目录，并在该目录下生成自定义算子安装包custom_opp_<target os>_<target architecture>.run。
```
bash build.sh
bash ./build_out/custom_opp_ubuntu_aarch64.run  
```

进入PyTorch接入工程，使用PyTorch调用方式调用AddCustom算子工程，并按照指导完成编译。

PytorchInvocation
├── op_plugin_patch         
├── run_op_plugin.sh      //  4.执行样例时，需要使用
└── test_ops_custom.py    //  步骤2启动工具时，需要使用

执行样例，样例执行过程中会自动生成测试数据，然后运行pytorch样例，最后检验运行结果。

bash run_op_plugin.sh
-- CMAKE_CCE_COMPILER: /usr/local/Ascend/ascend-toolkit/latest/toolkit/tools/ccec_compiler/bin/ccec
-- CMAKE_CURRENT_LIST_DIR: ${INSTALL_DIR}/AddKernelInvocation/cmake/Modules
-- ASCEND_PRODUCT_TYPE:
  Ascendxxxyy
-- ASCEND_CORE_TYPE:
  VectorCore
-- ASCEND_INSTALL_PATH:
  /usr/local/Ascend/ascend-toolkit/latest
-- The CXX compiler identification is GNU 10.3.1
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Configuring done
-- Generating done
-- Build files have been written to: ${INSTALL_DIR}/AddKernelInvocation/build
Scanning dependencies of target add_npu
[ 33%] Building CCE object cmake/npu/CMakeFiles/add_npu.dir/__/__/add_custom.cpp.o
[ 66%] Building CCE object cmake/npu/CMakeFiles/add_npu.dir/__/__/main.cpp.o
[100%] Linking CCE executable ../../../add_npu
[100%] Built target add_npu
${INSTALL_DIR}/AddKernelInvocation
INFO: compile op on ONBOARD succeed!
INFO: execute op on ONBOARD succeed!
test pass

手动导入算子调试信息。
- ${INSTALL_DIR}请替换为CANN软件安装后文件存储路径。例如，若安装的Ascend-cann-toolkit软件包，则安装后文件存储路径为：$HOME/Ascend/ascend-toolkit/latest。
- 在安装昇腾AI处理器的服务器执行npu-smi info命令进行查询，获取Chip Name信息。实际配置值为AscendChip Name，例如Chip Name取值为xxxyy，实际配置值为Ascendxxxyy。
```
(msdebug) export 
LAUNCH_KERNEL_PATH=${INSTALL_DIR}/opp/vendors/customize/op_impl/ai_core/tbe/kernel/SOC_VERSION/add_custom/AddCustom_1e04ee05ab491cc5ae9c3d5c9ee8950b.o
```

启动msdebug工具拉起Python程序，进入调试界面。

msdebug python3 test_ops_custom.py
(msdebug) target create "python3"
Current executable set to '/home/mindstudio/miniconda3/envs/py37/bin/python3' (aarch64).
(msdebug) settings set -- target.run-args  "test_ops_custom.py"
(msdebug)

设置断点。

根据指定源码文件与对应行号，在核函数中设置NPU断点。

如果host侧和kernel侧存在同名的算子实现文件，在设置断点时，推荐采用绝对路径进行设置，确保断点打在预期的文件上。
在对host侧源码文件进行打点时，可能会出现找不到实际位置的告警，类似如下提示：
```
(msdebug) b /home/xx/op_host/add_custom.cpp:24
Breakpoint 1: no locations (pending).
WARNING:  Unable to resolve breakpoint to any actual locations.
(msdebug)
```
在算子运行后，会自动找到实际位置，并自动设置断点。

(msdebug) b add_custom.cpp:60
Breakpoint 1: where = AddCustom_1e04ee05ab491cc5ae9c3d5c9ee8950b.o`::AddCustom_1e04ee05ab491cc5ae9c3d5c9ee8950b_1(uint8_t *, uint8_t *, uint8_t *, uint8_t *, uint8_t *) + 9912 [inlined] KernelAdd::Compute(int) + 3400 at add_custom.cpp:60:9, address = 0x00000000000026b8

运行程序，等待直到命中断点。

(msdebug) r
Process 197189 launched: '/home/miniconda3/envs/py38/bin/python3' (aarch64)
Process 197189 stopped and restarted: thread 1 received signal: SIGCHLD
...
[Launch of Kernel anonymous on Device 0]
Process 197189 stopped
[Switching to focus on Kernel anonymous, CoreId 8, Type aiv]
* thread #1, name = 'python3', stop reason = breakpoint 2.1
    frame #0: 0x00000000000026b8 AddCustom_1e04ee05ab491cc5ae9c3d5c9ee8950b.o`::AddCustom_1e04ee05ab491cc5ae9c3d5c9ee8950b_1(uint8_t *, uint8_t *, uint8_t *, uint8_t *, uint8_t *) [inlined] KernelAdd::Compute(this=0x000000000020efb8, progress=1) at add_custom.cpp:60:9
   57              LocalTensor<DTYPE_Y> yLocal = inQueueY.DeQue<DTYPE_Y>();
   58              LocalTensor<DTYPE_Z> zLocal = outQueueZ.AllocTensor<DTYPE_Z>();
   59              Add(zLocal, xLocal, yLocal, this->tileLength);
-> 60              outQueueZ.EnQue<DTYPE_Z>(zLocal);
   61              inQueueX.FreeTensor(xLocal);
   62              inQueueY.FreeTensor(yLocal);
   63          }
(msdebug)

其他调试操作可参考导入调试信息、打印内存与变量、msdebug调试信息展示及核切换等，与其操作一致。

删除断点，具体操作请参见删除行断点。

调试完以后，执行q命令并输入Y或y结束调试。

(msdebug) q
Quitting LLDB will kill one or more processes. Do you really want to proceed: [Y/n] y

调试PyTorch接口调用的算子

前提条件

操作步骤