Compile time logic for detecting streaming compatible and broadcast compatible data access patterns
A dataflow graph for an application has operation units that are configured to be producers and consumers of tensors. A write access pattern of a particular producer specifies an order in which the particular producer generates elements of a tensor, and a read access pattern of a corresponding consumer specifies an order in which the corresponding consumer processes the elements of the tensor. The technology disclosed detects conflicts between the producers and the corresponding consumers that have mismatches between the write access patterns and the read access patterns. A conflict occurs when the order in which the particular producer generates the elements of the tensor is different from the order in which the corresponding consumer processes the elements of the tensor. The technology disclosed resolves the conflicts by inserting buffers between the producers and the corresponding consumers.
1. A data processing system, comprising:
memory storing a dataflow graph for an application, the dataflow graph having operation units that are configured to be producers to produce tensors for execution of the application, and to be consumers to consume the tensors for execution of the application;
the memory storing write access patterns of the producers, and read access patterns of the consumers, wherein a write access pattern of a particular producer specifies an order in which the particular producer generates elements of a tensor, and a read access pattern of a corresponding consumer specifies an order in which the corresponding consumer processes the elements of the tensor; and
a compile time logic having access to the memory and configured to process the dataflow graph to
detect conflicts between certain ones of the producers and corresponding ones of the consumers that have mismatches between the write access patterns and the read access patterns, wherein a conflict occurs when the order in which the particular producer generates the elements of the tensor is different from the order in which the corresponding consumer processes the elements of the tensor; and
resolve the conflicts by inserting buffers between the certain ones of the producers and the corresponding ones of the consumers.
2. The data processing system of claim 1 , wherein the certain ones of the producers are configured to write elements of the tensors in the buffers in accordance with the write access patterns.
3. The data processing system of claim 2 , wherein the corresponding ones of the consumers are configured to read the elements of the tensors from the buffers in accordance with the read access patterns.
4. The data processing system of claim 1 , wherein the compile time logic is further configured to detect a group of corresponding consumers that have a same read access pattern for processing the elements of the tensor.
5. The data processing system of claim 4 , wherein the compile time logic is further configured to insert a single buffer between the group of corresponding consumers and the particular producer.
6. The data processing system of claim 5 , wherein the particular producer is configured to write the elements of the tensor in the single buffer in accordance with the write access pattern.
7. The data processing system of claim 6 , wherein consumers in the group of corresponding consumers are configured to read the elements of the tensor from the single buffer in accordance with the read access pattern.
8. The data processing system of claim 1 , wherein the write access patterns and the read access patterns are defined based on operation types implemented by the operation units.
9. The data processing system of claim 1 , wherein the compile time logic is further configured to generate a modified version of the dataflow graph with buffers inserted between the certain ones of the producers and the corresponding ones of the consumers.
10. The data processing system of claim 9 , wherein runtime logic is configured to allocate physical compute units and physical memory units of a reconfigurable processor to the modified version of the dataflow graph.
11. The data processing system of claim 10 , wherein the runtime logic is configured to execute the modified version of the dataflow graph on the reconfigurable processor based on the allocation.
12. A computer-implemented method, including:
storing a dataflow graph for an application, the dataflow graph having operation units that are configured to be producers to produce tensors for execution of the application, and to be consumers to consume the tensors for execution of the application;
storing write access patterns of the producers, and read access patterns of the consumers, wherein a write access pattern of a particular producer specifies an order in which the particular producer generates elements of a tensor, and a read access pattern of a corresponding consumer specifies an order in which the corresponding consumer processes the elements of the tensor; and
detecting conflicts between certain ones of the producers and corresponding ones of the consumers that have mismatches between the write access patterns and the read access patterns, wherein a conflict occurs when the order in which the particular producer generates the elements of the tensor is different from the order in which the corresponding consumer processes the elements of the tensor; and
resolving the conflicts by inserting buffers between the certain ones of the producers and the corresponding ones of the consumers.
13. The computer-implemented method of claim 12 , wherein the certain ones of the producers are configured to write elements of the tensors in the buffers in accordance with the write access patterns.
14. The computer-implemented method of claim 13 , wherein the corresponding ones of the consumers are configured to read the elements of the tensors from the buffers in accordance with the read access patterns.
15. The computer-implemented method of claim 12 , wherein the compile time logic is further configured to detect a group of corresponding consumers that have a same read access pattern for processing the elements of the tensor.
16. The computer-implemented method of claim 15 , wherein the compile time logic is further configured to insert a single buffer between the group of corresponding consumers and the particular producer.
17. The computer-implemented method of claim 16 , wherein the particular producer is configured to write the elements of the tensor in the single buffer in accordance with the write access pattern.
18. A non-transitory computer readable storage medium impressed with computer program instructions to process data, the instructions, when executed on a processor, implement a method comprising:
storing a dataflow graph for an application, the dataflow graph having operation units that are configured to be producers to produce tensors for execution of the application, and to be consumers to consume the tensors for execution of the application;
storing write access patterns of the producers, and read access patterns of the consumers, wherein a write access pattern of a particular producer specifies an order in which the particular producer generates elements of a tensor, and a read access pattern of a corresponding consumer specifies an order in which the corresponding consumer processes the elements of the tensor; and
detecting conflicts between certain ones of the producers and corresponding ones of the consumers that have mismatches between the write access patterns and the read access patterns, wherein a conflict occurs when the order in which the particular producer generates the elements of the tensor is different from the order in which the corresponding consumer processes the elements of the tensor; and
resolving the conflicts by inserting buffers between the certain ones of the producers and the corresponding ones of the consumers.
19. The non-transitory computer readable storage medium of claim 18 , wherein the certain ones of the producers are configured to write elements of the tensors in the buffers in accordance with the write access patterns.
20. The non-transitory computer readable storage medium of claim 19 , wherein the corresponding ones of the consumers are configured to read the elements of the tensors from the buffers in accordance with the read access patterns.