Compilation method

Patent metadata

Field	Value
Publication number	US-10691432-B1
Application number	US-201916277178-A
Country	US
Kind code	B1
Filing date	Feb 15, 2019
Priority date	Jan 3, 2019
Publication date	Jun 23, 2020
Grant date	Jun 23, 2020

Abstract

A method for generating a program to run on multiple tiles. The method comprises: receiving an input graph comprising data nodes, compute vertices and edges; receiving an initial tile-mapping specifying which data nodes and vertices are allocated to which tile; and determining a subgraph of the input graph that meets one or more heuristic rules. The rules comprises: the subgraph comprises at least one data node, the subgraph spans no more than a threshold number of tiles in the initial tile-mapping, and the subgraph comprises at least a minimum number of edges outputting to one or more vertices on one or more other tiles. The method further comprises adapting the initial mapping to migrate the data nodes and any vertices of the determined subgraph to said one or more other tiles, and compiling the executable program from the graph with the vertices and data nodes allocated by the adapted mapping.

First claim

What is claimed is:

1. A computer-implemented method for generating an executable program to run on a processing system comprising one or more chips each comprising a plurality of tiles, each tile comprising a respective processing unit and memory; the method comprising:
receiving an input graph comprising a plurality of data nodes, a plurality of compute vertices and a plurality of directional edges, each edge representing an output from a data node input to a compute vertex or an output from a compute vertex input to a data node, each data node representing a variable and/or constant, and each compute vertex representing one or more computations to perform on the input to the compute vertex in order to result in the output from that compute vertex;
receiving an initial tile-mapping specifying which of the data nodes and vertices are allocated to be run on which of the tiles;
determining a subgraph of the input graph that meets one or more heuristic rules, the rules comprising:
the subgraph comprises at least one data node,
the subgraph spans no more than a threshold number of tiles in the initial tile-mapping, and
the subgraph comprises at least a minimum number of edges outputting to one or more vertices on one or more others of the tiles;

adapting the initial mapping to migrate the data nodes and any vertices of the determined subgraph to said one or more other tiles; and
compiling the executable program from the graph with the vertices and data nodes configured to run on the tiles specified by the adapted mapping.

2. The method of claim 1 , wherein the heuristic rules further comprise:
all data nodes in the subgraph are scalars.

3. The method of claim 1 , wherein the heuristic rules further comprise:
any compute vertex in the subgraph outputs to at least one data node in the subgraph.

4. The method of claim 1 , wherein said threshold is one.

5. The method of claim 1 , wherein said threshold is the same for both data nodes and compute vertices.

6. The method of claim 1 , wherein said rule that the subgraph spans no more than a threshold number of tiles comprises:
the data nodes of the subgraph span no more than a first threshold number of tiles, and
the vertices of the subgraph span no more than a second threshold number of tiles wherein the second threshold is different than the first.

7. The method of claim 6 , wherein the first threshold is one and the second threshold is greater than one, or the second threshold is one and the first threshold is greater than one.

8. The method of claim 1 , wherein said minimum number is one.

9. The method of claim 1 , wherein said minimum number is greater than said threshold.

10. The method of claim 1 , wherein the heuristic rules further comprise:
any compute vertex in the subgraph is allocated to the same tile as at least one data node in the subgraph in the initial mapping.

11. The method of claim 1 , wherein the vertices are divided amongst a plurality of compute sets ordered according to an order of execution; and wherein said rules further comprise:
the vertices in the subgraph are all in the same compute set.

12. The method of claim 1 , wherein the determined subgraph comprises edges outputting to compute vertices on a multiple others of the tiles, and the migration comprises replicating the determined subgraph across each of the multiple tiles.

13. The method of claim 1 , wherein said determination of the subgraph comprises performing a search comprising:
selecting one of the data nodes as a starting point; and
performing a search comprising expanding a candidate subgraph from the starting point and terminating the expansion of the candidate subgraph, or a branch thereof, upon encountering a node, vertex or edge that fails to match one or more search criteria.

14. The method of claim 13 , wherein the search criteria comprise at least one or more of said rules that:
the subgraph spans no more than a threshold number of tiles in the initial tile-mapping,
the data nodes of the subgraph span no more than a first threshold number of tiles,
the vertices of the subgraph span no more than a second threshold number of tiles wherein the second threshold is different than the first,
all data nodes in the subgraph are scalars,
any compute vertex in the subgraph is allocated to the same tile as at least one data node in the subgraph in the initial mapping, and/or
the vertices in the subgraph are all in the same compute set.

15. The method of claim 13 , wherein said search criteria comprise at least a condition that the subgraph has not exceeded a maximum search depth.

16. The method of claim 13 , where said search is a breadth first search.

17. The method of claim 13 , wherein said search is a depth first search.

18. The method of claim 13 , comprising performing a plurality of iterations of said search expanding from different starting points until finding a candidate subgraph that, after said termination, meets one or more filtering criteria; and selecting the found subgraph as the determined subgraph to use for said adaptation.

19. The method of claim 18 , wherein said filtering criteria comprise one or more of said rules that:
the subgraph comprises at least one data node,
the subgraph comprises multiple edges outputting to one or more vertices on one or more others of the processor modules,
any compute vertex in the subgraph is allocated to the same tile as at least one data node in the subgraph in the initial mapping, and/or
the vertices in the subgraph are all in the same compute set.

20. The method of claim 13 , comprising applying a restriction that the or each starting point must be a constant data node.

21. The method of claim 13 , comprising applying a restriction that the or each starting point must be a scalar data node.

22. The method of claim 1 , wherein said compilation further comprises configuring the program to synchronize between the tiles according to a bulk synchronous parallel scheme.

23. The method of claim 1 , wherein said graph comprises a neural network, and the program comprises an algorithm configured to perform machine learning using the neural network.

24. The method of claim 1 , wherein at least some of said data nodes represent tensors.

25. A computer comprising a processor and memory, the memory storing a software tool, the software tool comprising software configured so as when run on said computer, causes the computer to perform a method comprising:
receiving an input graph comprising a plurality of data nodes, a plurality of compute vertices and a plurality of directional edges, each edge representing an output from a data node input to a compute vertex or an output from a compute vertex input to a data node, each data node representing a variable and/or constant, and each compute vertex representing one or more computations to perform on the input to the compute vertex in order to result in the output from that compute vertex;
receiving an initial tile-mapping allocating the data nodes and the vertices to a plurality of tiles in a processing system;
determining a subgraph of the input graph that meets the following rules: a) the subgraph comprises a subset of the nodes and has at least one data node of the plurality of data nodes, b) the subgraph spans no more than a threshold number of tiles in the initial tile-mapping, and c) the subgraph comprises at least a minimum number of edges outputting to o

Assignees

• Graphcore Ltd

Inventors

• Pupilli Mark Lloyd
• Lacey David

Classifications

• G06F16/9024

  Graphs; Linked lists (G06F16/9027 takes precedence) · CPC title

• G06F8/451 Primary

  Code distribution (considering CPU load at run-time G06F9/505; load rebalancing G06F9/5083) · CPC title

• G06F8/10 Primary

  Requirements analysis; Specification techniques · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title

• G06N3/063

  using electronic means · CPC title