Operation fusion in nested meta-pipeline loops
A method for improving throughput in a reconfigurable computing system includes detecting, in an algebraic representation of a computing task for a reconfigurable dataflow processor, an outer meta-pipeline loop, detecting an inner meta-pipeline loop nested within the outer meta-pipeline loop, and determining that the inner meta-pipeline loop and the outer meta-pipeline loop each conduct a common operation. The method also includes fusing the common operation for the inner meta-pipeline loop and the outer meta-pipeline loop into a single operation within the inner meta-pipeline loop. The instances of the common operation may be fused if the output of a first instance of the common operation is the source for a second instance of the common operation. Examples of the common operation include an accumulator operation, a re-read operation, and a temporal (chip buffer synchronized) operation such as a temporal concatenation operation and a temporal slicing operation.
1. A system for improving throughput in reconfigurable dataflow processors, the system comprising:
a host computer comprising an optimization module and configured to conduct a method of transforming a high-level program of a computing task into configuration data executable by the reconfigurable dataflow processor (RDP) including one or more arrays of configurable units, the method comprising:
detecting, in an algebraic representation of the computing task for the RDP, an outer meta-pipeline loop;
detecting, in the algebraic representation, an inner meta-pipeline loop nested within the outer meta-pipeline loop;
determining that the inner meta-pipeline loop and the outer meta-pipeline loop each conduct a common operation; and
fusing the common operation for the inner meta-pipeline loop and the outer meta-pipeline loop into a single operation within the inner meta-pipeline loop;
generating the configuration data of the computing task for the RDP processor including placement and routing of configurable units, wherein the configuration data, when loaded onto an instance of the one or more arrays of configurable units of the RDP processor, causes the one or more arrays of configurable units to implement at least the computing task; and
storing the configuration data in a non-transitory computer-readable storage medium.
2. The system of claim 1 , wherein the output of a first instance of the common operation is the source for a second instance of the common operation.
3. The system of claim 1 , wherein the common operation is one of an accumulator operation, a re-read operation and a temporal operation.
4. The system of claim 1 , wherein the algebraic representation comprises a compute graph.
5. The system of claim 4 , wherein nodes in the compute graph correspond to tensor operations.
6. The system of claim 1 , wherein the algebraic representation comprises code blocks.
7. The system of claim 6 , wherein statements in the code blocks correspond to tensor operations.
8. The system of claim 1 , wherein the host computer further comprises an allocation module configured to allocate configurable units based on the configuration data.
9. The system of claim 8 , the host computer further comprises a place and route module configured to place and route the configurable units based on the configuration data.
10. A method for improving throughput in a reconfigurable computing system, the method configured to transform a high-level program of a computing task into configuration data executable by a reconfigurable dataflow processor (RDP) including one or more arrays of configurable units and comprising:
detecting, in an algebraic representation of a computing task for the RDP, an outer meta-pipeline loop;
detecting, in the algebraic representation, an inner meta-pipeline loop nested within the outer meta-pipeline loop;
determining that the inner meta-pipeline loop and the outer meta-pipeline loop each conduct a common operation; and
fusing the common operation for the inner meta-pipeline loop and the outer meta-pipeline loop into a single operation within the inner meta-pipeline loop;
generating the configuration data of the computing task for the RDP processor including placement and routing of configurable units, wherein the configuration data, when loaded onto an instance of the one or more arrays of configurable units of the RDP processor, causes the one or more arrays of configurable units to implement at least the computing task; and
storing the configuration data in a non-transitory computer-readable storage medium.
11. The method of claim 10 , wherein the output of a first instance of the common operation is the source for a second instance of the common operation.
12. The method of claim 10 , wherein the common operation is one of an accumulator operation, a re-read operation and a temporal operation.
13. The method of claim 10 , wherein the algebraic representation comprises a compute graph.
14. The method of claim 13 , wherein nodes in the compute graph correspond to tensor operations.
15. The method of claim 10 , wherein the algebraic representation comprises code blocks.
16. The method of claim 15 , wherein statements in the code blocks correspond to tensor operations.
17. The method of claim 10 , further comprising allocating, placing and routing configurable units and connections based on the configuration data.
18. The method of claim 17 , wherein the configurable units comprise one or more of memory units, compute units and switches.
19. The method of claim 10 , further comprising configuring the RDP to perform the computing task and performing the computing task.
20. A computer program product comprising a computer readable storage medium having program instructions embodied therewith, wherein the computer readable storage medium is not a transitory signal per se, wherein the program instructions are executable by a processor to cause the processor to conduct a method of transforming a high-level program of a computing task into configuration data executable by a reconfigurable dataflow processor (RDP) including one or more arrays of configurable units, the method comprising:
detecting, in an algebraic representation of a computing task for the RDP, an outer meta-pipeline loop;
detecting, in the algebraic representation, an inner meta-pipeline loop nested within the outer meta-pipeline loop;
determining that the inner meta-pipeline loop and the outer meta-pipeline loop each conduct a common operation; and
fusing the common operation for the inner meta-pipeline loop and the outer meta-pipeline loop into a single operation within the inner meta-pipeline loop;
generating the configuration data of the computing task for the RDP processor including placement and routing of configurable units, wherein the configuration data, when loaded onto an instance of the one or more arrays of configurable units of the RDP processor, causes the one or more arrays of configurable units to implement at least the computing task; and
storing the configuration data in a non-transitory computer-readable storage medium.