Subgraph tile fusion in a convolutional neural network

1. A method of subgraph tile fusion in a convolutional neural network, comprising:

partitioning a network into at least one subgraph node;

determining a layer order of at least one layer of the at least one subgraph node;

determining a input layer of the at least one subgraph node;

determining a weight layer of the at least one subgraph node;

determining a output layer of the at least one subgraph node; and

fusing the at least one subgraph node, the input layer, the weight layer and the output layer in the layer order.

2. The method of subgraph tile fusion of claim 1 , further comprising controlling the fused at least one subgraph node control from within the fused at least one subgraph node.

3. The method of subgraph tile fusion of claim 1 , further comprising receiving input from an input buffer table.

4. The method of subgraph tile fusion of claim 3 , further comprising partitioning the input buffer table into a plurality of input buffer tensor tiles.

5. The method of subgraph tile fusion of claim 1 , further comprising sending output from an output buffer table.

6. The method of subgraph tile fusion of claim 5 , further comprising partitioning the output buffer table into a plurality of output buffer tensor tiles.

7. The method of subgraph tile fusion of claim 1 , further comprising receiving weight data from a weight buffer table.

8. The method of subgraph tile fusion of claim 7 , further comprising partitioning the weight buffer table into a plurality of weight buffer tensor tiles.

9. The method of subgraph tile fusion of claim 1 , further comprising executing the at least one subgraph node in the layer order.

10. The method of subgraph tile fusion of claim 1 , further comprising adding an additional layer having two input layers.

11. The method of subgraph tile fusion of claim 1 , further comprising retaining the input layer based on the input.

12. The method of subgraph tile fusion of claim 1 , further comprising linking the output layer to a set of multiple nodes.

13. A method of subgraph tile fusion in a convolutional neural network, comprising:

partitioning a network operation into at least one subgraph node;

determining a layer order of at least one layer of the subgraph node;

determining a input layer of the subgraph node;

determining a weight layer of the subgraph node;

determining a output layer of the subgraph node; and

fusing the at least one subgraph node, the input layer, the weight layer and the output layer in the layer order.

14. The method of subgraph tile fusion of claim 13 , further comprising:

receiving input from an input buffer table;

receiving weight data from a weight buffer table; and

sending output from an output buffer table.

15. The method of subgraph tile fusion of claim 14 , further comprising:

partitioning the input buffer table into a plurality of input buffer tensor tiles;

partitioning the output buffer table into a plurality of output buffer tensor tiles; and

partitioning the weight buffer table into a plurality of weight buffer tensor tiles.

16. The method of subgraph tile fusion of claim 14 , further comprising:

reading one input from the input buffer table; and

outputting one output from the output buffer table.

17. The method of subgraph tile fusion of claim 13 , further comprising executing the at least one subgraph node in the layer order.

18. The method of subgraph tile fusion of claim 13 , further comprising adding an additional layer having two input layers.

19. The method of subgraph tile fusion of claim 13 , further comprising retaining the input layer based on the input.

20. The method of subgraph tile fusion of claim 13 , further comprising linking the output layer to a set of multiple nodes.