Ascend 算子开发实战+基础算子开发命题笔记（以ThreeNN为例）

1. 编写构建算子工程的JSON文件

2. 

3. [
       {
           "op": "ThreeNNCustom",
           "language":"cpp",
           "input_desc": [
               {
                   "name": "xyz1",
                   "param_type": "required",
                   "format": [
                       "ND"
                   ],
                   "type": [
                       "fp32"
                   ]
               },
               {
                   "name": "xyz2",
                   "param_type": "required",
                   "format": [
                       "ND"
                   ],
                   "type": [
                       "fp32"
                   ]
               }
           ],
           "output_desc": [
               {
                   "name": "dist",
                   "param_type": "required",
                   "format": [
                       "ND"
                   ],
                   "type": [
                       "fp32"
                   ]
               },
               {
                   "name": "idx",
                   "param_type": "required",
                   "format": [
                       "ND"
                   ],
                   "type": [
                       "int32"
                   ]
               }
           ]
       }
   ]

4. 将/usr/local/Ascend/ascend-toolkit/8.0.RC2.alpha002/python/site-packages/bin加入PATH

5. 通过以下命令新建工程 msopgen gen -i ${JSON文件} -c ai_core-ascend310b -lan cpp -out {工程初始化目录}

6. 新建完后，可能要调整CMakePresets.json文件，设置正确的ascend package 地址

   "ASCEND_CANN_PACKAGE_PATH": {
       "type": "PATH",
       "value": "/usr/local/Ascend/ascend-toolkit/latest"
   },

7. 运行build.sh，编译算子，完成后再运行build_out/custom_opp_ubuntu_aarch64.run将自己写的算子安装到本地库里面

8. 从下载测试样例，解压缩后，进入ThreeNNCase/ThreeNNCase1文件夹

9. src/oprunner.cpp文件可能需要的修改如下

   #include "aclnn_three_nn.h" -> #include "aclnn_three_nn_custom.h"
   
   auto ret = aclnnThreeNNGetWorkspaceSize(inputTensor_[0], inputTensor_[1], outputTensor_[0],  outputTensor_[1], &workspaceSize, &handle);
   ->
   auto ret = aclnnThreeNNCustomGetWorkspaceSize(inputTensor_[0], inputTensor_[1], outputTensor_[0],  outputTensor_[1], &workspaceSize, &handle);
   
   ret = aclnnThreeNN(workspace, workspaceSize, handle, stream);
   ->
   ret = aclnnThreeNNCustom(workspace, workspaceSize, handle, stream);

10. 运行run.sh进行评测