阅读 149

TVM性能评估分析(七)

TVM性能评估分析(七)

TVM性能评估分析(七)

 

 

 Figure 1.  Performance Improvement

 

 

 Figure 2.  Depthwise convolution

 

 

Figure 3.  Data Fusion

 

 

 Figure 4.  Data Fusion(2)

 

 

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

 

 

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

 

 

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

 

 

 Figure 8.  Computational Graph

 

 

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

 

 

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

 

 

 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.

 

 

 Figure 12.  Multi-language and Platform Support

 

 

 Figure 13.  Remote Deployment and Execution

 

 

 Table 1.  Raspberry Pi

 

 

 Figure 14.  GPU Results


文章分类
后端
文章标签
版权声明:本站是系统测试站点,无实际运营。本文内容由互联网用户自发贡献,该文观点仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌抄袭侵权/违法违规的内容, 请发送邮件至 XXXXXXo@163.com 举报,一经查实,本站将立刻删除。
相关推荐