当前位置:网站首页>[cann training camp] AI CPU operator development based on shengteng cann platform
[cann training camp] AI CPU operator development based on shengteng cann platform
2022-07-21 03:36:00 【Tianyi Li 1997】
Preface
Describes how to develop CANN AICPU Custom operators , With a AICPU Operator as an example , analysis + Realization + Test and verify an end-to-end complete development process .
summary
AI CPU operator , Is running in shengteng AI In the processor AI CPU The operation of expressing a complete computational logic on a computing unit , In the following cases , Developers need to customize AI CPU operator .
stay NN In the process of model training or reasoning , Transform the third-party open source framework into adaptive shengteng AI The processor model encountered rising AI Operators not supported by the processor . here , In order to quickly get through the model execution process , Users can customize AI CPU Operator for function adjustment , Improve the debugging efficiency . After the function is adjusted , In the subsequent performance testing process AI CPU The custom operator is converted to TBE Operator implementation .
In some cases , Can't be realized in AI Core Custom operators running on ( For example, some operators need Complex32、Complex64 type , but AI Core Command does not support ; Another example is the operator that contains a large number of scalar calculations , and AI Core Not good at scalar processing ), At this time, you can develop AI CPU Custom operator to achieve ascension AI Processor support for this operator .
The goal is
Can be based on AI CPU People who develop simple operators , It can be achieved :
Study AI CPU Basic implementation principle and method of operator .
Can be based on examples in the course , Extension for other customization AI CPU Operator development .
If you have the following skills , Can better complete learning :
Have C++ Program development ability
Understand mathematical expressions
Machine learning 、 Deep learning has a certain understanding
understand Ascend Platform operation process and principle
understand Ascend Platform TBE Custom operator development process
The basic concept of operator
Ascend 310 Processor architecture logic
AI CPU Operator development process
Operator analysis
Introduction to directory structure
边栏推荐
- Load different plist files according to different configurations
- 达梦dem部署
- Swift Archiving序列化
- 达梦AWR报告提取
- Centos8(linux)安装mysql8.0.28
- (9) Pytorch deep learning: convolutional neural network (the inception module in googlenet network architecture is this convolutional neural network architecture)
- MIMO-OFDM无线通信技术及MATLAB实现(2)-SISO下的室外信道模型
- Opencv series of tutorials (I): opencv reads pictures and videos in specified folders and calls cameras
- Redis持久化
- (7) Pytorch deep learning: full connection layer network
猜你喜欢
基于弹性云服务器的昇腾AI应用开发随笔【与云原生的故事】
Li Hongyi 2020 machine learning -- P11 logistic progression
玩转CANN目标检测与识别一站式方案【介绍篇】
MIMO-OFDM无线通信技术及MATLAB实现(2)-SISO下的室外信道模型
Netcat simple gadget simulates client / server
(一)PyTorch深度学习:线性模型训练
(五)PyTorch深度学习:Logistic回归
Redis持久化
Centos8(linux)安装mysql8.0.28
Paper reading -- risk resistant resource allocation for embB and urllc coexistence under m/g/1 queuing model
随机推荐
MIMO - OFDM Wireless Communication Technology and matlab Implementation (2) - outdoor Channel Model under SISO
Task scheduling: common types and tools
数仓 -OLAP OLTP 建模方法
Swift Archiving序列化
基于点云的深度学习方法综述
Redis主从复制
Basic overview of data warehouse
李宏毅2020机器学习笔记---P10 Classification
达梦索引管理
Centos8(linux)安装mysql8.0.28
Flutter 小结
(四)PyTorch深度学习:PytTorch实现线性回归
Centos8 (Linux) install mysql8.0.28
Introduction to microservice theory
达梦免密登录
Li Hongyi machine learning 2020---p12 brief introduction of DL & p15 why DL
如何使用 IDEA 打 jar 包
Dameng collects statistics
Basic architecture of data warehouse -- layered functions
Spark SQL 中的小 TIPS