diff --git "a/AICore Profiling\345\267\245\345\205\267\344\275\277\347\224\250\346\214\207\345\257\274\344\271\246.md" "b/AICore Profiling\345\267\245\345\205\267\344\275\277\347\224\250\346\214\207\345\257\274\344\271\246.md"
new file mode 100644
index 0000000000000000000000000000000000000000..75f01c8e3161afc4a7982e5a1f25dce9db92ecf4
--- /dev/null
+++ "b/AICore Profiling\345\267\245\345\205\267\344\275\277\347\224\250\346\214\207\345\257\274\344\271\246.md"	
@@ -0,0 +1,67 @@
+## 1 背景  
+（1）模型调测过程中，性能分析是一大块工作，所以一款好用的profiling工具必不可少。  
+（2）630之前的profiling工具其实就已经有了，但易用性上直接让人感到反感，这次改版之后，改进很大，数据采集和数据解析在易用性上也有很大提升。  
+
+## 2 数据采集（训练任务JOB）  
+其他方式不在这里介绍，这里只介绍设置环境变量的方式，开启profiling功能：  
+```
+export PROFILING_MODE=true
+export PROFILING_OPTIONS='{"output": "./cann_profiling", "training_trace": "on", "task_trace": "on", "aicpu": "on", "fp_point": "", "bp_point": "", "aic_metrics": "PipeUtilization"}'
+```
+其中，output指明采集数据的存放目录，需提前创建好目录, aicpu为aicpu算子开关，aic_metrics为aicore算子开关，其他参数见说明[CANN V100R020C10 开发辅助工具指南 (训练) 01](https://support.huawei.com/enterprise/zh/doc/EDOC1100164832/6f4033fd)  
+
+设置好环境变量后，然后执行训练，之后会在./cann_profiling目录下产生profiling数据：  
+![](https://gitee.com/zwx5317131/ascend-pytorch-crowdintelligence-doc/raw/master/figures/aicore_profiling_fig1.png)
+
+
+## 3 数据解析
+### 3.1 解析工具
+解析数据必须要用到昇腾软件包的安装目录下的msprof工具，一般默认路径为：
+```
+x86路径为:  
+/usr/local/Ascend/ascend-toolkit/latest/x86_64-linux/toolkit/tools/profiler/profiler_tool/analysis/msprof/msprof.py   
+
+arm路径为：  
+/usr/local/Ascend/ascend-toolkit/latest/aarch64-linux/toolkit/tools/profiler/profiler_tool/analysis/msprof/msprof.py 
+```  
+以下指令以x86_64环境为例
+
+### 3.2 解析profiling数据
+```
+python3.7 /usr/local/Ascend/ascend-toolkit/latest/x86_64-linux/toolkit/tools/profiler/profiler_tool/analysis/msprof/msprof.py import -dir ./cann_profiling/
+```
+其中，./cann_profiling为上面采集数据的保存路径
+执行过程：
+![](https://gitee.com/zwx5317131/ascend-pytorch-crowdintelligence-doc/raw/master/figures/aicore_profiling_fig2.png)
+
+执行之后会在cann_profiling目录下生成sqlite等数据目录：  
+![](https://gitee.com/zwx5317131/ascend-pytorch-crowdintelligence-doc/raw/master/figures/aicore_profiling_fig3.png)
+
+### 3.3 导出timeline数据
+导出timeline数据，执行如下命令： 
+```
+python3.7 /usr/local/Ascend/ascend-toolkit/latest/x86_64-linux/toolkit/tools/profiler/profiler_tool/analysis/msprof/msprof.py export timeline -dir ./cann_profiling/
+```
+如果需要指定某一个迭代step，可设置参数：-iteration-id  
+执行之后会生成timeline文件夹：  
+![](https://gitee.com/zwx5317131/ascend-pytorch-crowdintelligence-doc/raw/master/figures/aicore_profiling_fig4.png)
+
+里面的json文件可以用chrome://tracing查看  
+![](https://gitee.com/zwx5317131/ascend-pytorch-crowdintelligence-doc/raw/master/figures/aicore_profiling_fig5.png)
+
+### 3.4 导出summary数据  
+导出summary数据，执行如下命令：  
+```
+python3.7 /usr/local/Ascend/ascend-toolkit/latest/x86_64-linux/toolkit/tools/profiler/profiler_tool/analysis/msprof/msprof.py export summary -dir ./cann_profiling/
+```
+
+如果需要指定某一个迭代step，可设置参数：-iteration-id  
+执行之后会生成summary文件夹，里面csv文件就是summary数据，能看到算子名称，算子执行顺序，算子耗时。 
+![](https://gitee.com/zwx5317131/ascend-pytorch-crowdintelligence-doc/raw/master/figures/aicore_profiling_fig6.png) 
+
+csv文件  
+![](https://gitee.com/zwx5317131/ascend-pytorch-crowdintelligence-doc/raw/master/figures/aicore_profiling_fig7.png)
+
+## 4 展望
+（1）该工具还可以profiling系统性能数据，如PCIE，DVPP，HBM等。  
+（2）目前算子的input shape及其dtype、format，output shape及其dtype、format等信息（task_info信息）pytorch场景没有生成，tf场景是有的。因为对于op_based场景，GE没有上报这些信息。已经提了这个需求。  
diff --git a/figures/aicore_profiling_fig1.png b/figures/aicore_profiling_fig1.png
new file mode 100644
index 0000000000000000000000000000000000000000..cccb56eab432c1ac8c363bfc3d48e582c4f4d67c
Binary files /dev/null and b/figures/aicore_profiling_fig1.png differ
diff --git a/figures/aicore_profiling_fig2.png b/figures/aicore_profiling_fig2.png
new file mode 100644
index 0000000000000000000000000000000000000000..f983683ff7fd2e15d2081ecdc8ea2410508e625a
Binary files /dev/null and b/figures/aicore_profiling_fig2.png differ
diff --git a/figures/aicore_profiling_fig3.png b/figures/aicore_profiling_fig3.png
new file mode 100644
index 0000000000000000000000000000000000000000..5e4ae0c5b57d36b334e168e182fb5e19aabc3251
Binary files /dev/null and b/figures/aicore_profiling_fig3.png differ
diff --git a/figures/aicore_profiling_fig4.png b/figures/aicore_profiling_fig4.png
new file mode 100644
index 0000000000000000000000000000000000000000..2192c7959db28c963dc69bd7c623e683ae02b69f
Binary files /dev/null and b/figures/aicore_profiling_fig4.png differ
diff --git a/figures/aicore_profiling_fig5.png b/figures/aicore_profiling_fig5.png
new file mode 100644
index 0000000000000000000000000000000000000000..9470f58b866d7e75f1adcaf230e55b228174d90b
Binary files /dev/null and b/figures/aicore_profiling_fig5.png differ
diff --git a/figures/aicore_profiling_fig6.png b/figures/aicore_profiling_fig6.png
new file mode 100644
index 0000000000000000000000000000000000000000..65440a7b8763b801a0d7da66022b5785e8dfd7d2
Binary files /dev/null and b/figures/aicore_profiling_fig6.png differ
diff --git a/figures/aicore_profiling_fig7.png b/figures/aicore_profiling_fig7.png
new file mode 100644
index 0000000000000000000000000000000000000000..f9f9bae08ef6674dfbba1738c3cd797ea98a39ac
Binary files /dev/null and b/figures/aicore_profiling_fig7.png differ