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