标签：compile end Figure workflow 性能 TVM device 评估

TVM性能评估分析（六）

 

 

 Figure 1.  The workflow of development PC, compile, deploy to the device, test, then modify the codes again to see whether it accelerates.

 

 

 Figure 2.   The Android APP takes shared library as input and runs compiled functions on the mobile phone. 

 

 

 Figure 3.  Build TVM functions and NDArrays on a remote device. The ability to cross-compile to different platforms makes it easy to develop on one platform and test on another.

 

 

 Figure 4.  The instruction to build for your Android device. Once the APK is built, sign it using apps/android_rpc/dev_tools and install it on the phone. 

 

 

 Figure 5.  The NNVM compiler support of TVM stack, we can now directly compile descriptions from deep learning frameworks and compile them to bare metal code that runs on AMD GPUs.

 

 

 Figure 6.  With ROCm backend, the generic workflow 

 

 

 Figure 7.   The ONNX library to load the ONNX model into the Protocol buffer object. 

 

 

 Figure 8.  An end to end compilation pipeline from front-end deep learning frameworks to bare metal hardwares.

 

Figure 9.  Typical workflow of NNVM Compiler

 

 

 Figure 10.  Separation of Optimization and Deployment

 

 

 Figure 11.  Time Cost of Inference on K80

 

 

 Figure 12.  The cost of inference on Raspberry PI

标签：compile,end,Figure,workflow,性能,TVM,device,评估
来源： https://www.cnblogs.com/wujianming-110117/p/14826997.html

本站声明： 1. iCode9 技术分享网（下文简称本站）提供的所有内容，仅供技术学习、探讨和分享；
2. 关于本站的所有留言、评论、转载及引用，纯属内容发起人的个人观点，与本站观点和立场无关；
3. 关于本站的所有言论和文字，纯属内容发起人的个人观点，与本站观点和立场无关；
4. 本站文章均是网友提供，不完全保证技术分享内容的完整性、准确性、时效性、风险性和版权归属；如您发现该文章侵犯了您的权益，可联系我们第一时间进行删除；
5. 本站为非盈利性的个人网站，所有内容不会用来进行牟利，也不会利用任何形式的广告来间接获益，纯粹是为了广大技术爱好者提供技术内容和技术思想的分享性交流网站。