码迷,mamicode.com
首页 > 其他好文 > 详细

TVM性能评估分析(七)

时间:2021-06-02 16:46:26      阅读:0      评论:0      收藏:0      [点我收藏+]

标签:gpu   ati   ase   ast   signed   exe   bank   loading   inf   

TVM性能评估分析(七)

 技术图片

 

 Figure 1.  Performance Improvement

 技术图片

 

 Figure 2.  Depthwise convolution

 技术图片

 

Figure 3.  Data Fusion

 技术图片

 

 Figure 4.  Data Fusion(2)

 技术图片

 

 Figure 5.  Shared memory can be seen as cache in GPU. It is on-chip and much faster than global memory.

 技术图片

 

 Figure 6.   Shared memory banks are organized such that successive addresses are assigned to successive banks. 

 技术图片

 

 Figure 7.  Consecutive threads access consecutive memory addresses, thus avoiding bank conflicts

 技术图片

 

 Figure 8.  Computational Graph

 技术图片

 

 Figure 9.  Sublinear memory optimization functionality that allows user to train 1000 layers of ImageNet ResNet on a single GPU.

 技术图片

 

 Figure 10.  We build a low level representation which is based on index formula, with additional support for recurrence computation.

 技术图片

 

 Figure 11.  The algorithms described in TVM are then processed in a scheduling phase to apply transformations that are tailored to the target hardware back-end.

 技术图片

 

 Figure 12.  Multi-language and Platform Support

 技术图片

 

 Figure 13.  Remote Deployment and Execution

 技术图片

 

 Table 1.  Raspberry Pi

 技术图片

 

 Figure 14.  GPU Results

 

TVM性能评估分析(七)

标签:gpu   ati   ase   ast   signed   exe   bank   loading   inf   

原文地址:https://www.cnblogs.com/wujianming-110117/p/14827032.html

(0)
(0)
   
举报
评论 一句话评论(0
登录后才能评论!
© 2014 mamicode.com 版权所有  联系我们:gaon5@hotmail.com
迷上了代码!