Compile time logic for inserting a buffer between a producer operation unit and a consumer operation unit in a dataflow graph

A computer-implemented method, including:

• obtaining dataflow graph for an application;
• determining a write access pattern for elements of a tensor generated by a producer operation unit of the dataflow graph for an application;
• determining a read access pattern for the elements of the tensor processed by consumer operation unit of the dataflow graph;
• identifying a group of consumer operation units having a same read access pattern,
• comparing the write access pattern to the read access pattern;
• classifying a data access pattern between the producer operation unit and the group of consumer operation units as streaming compatible if the read access pattern is same as the write access pattern,
• classifying the data access pattern between the producer operation unit and the group of consumer operation units as broadcast compatible if the read access pattern is different from the write access pattern, and
• modifying the dataflow graph by inserting only a single buffer between the producer operation unit and the group of consumer operation units if the data access pattern is broadcast compatible to create a modified dataflow graph, thereby resolving the difference between the write access pattern and the read access pattern using the single buffer.

A non-transitory computer readable storage medium storing computer program instructions to process data, the instructions, when executed on a processor, implement a method comprising:

• obtaining dataflow graph for an application;
• determining a write access pattern for elements of a tensor generated by a producer operation unit of the dataflow graph for an application;
• determining a read access pattern for the elements of the tensor processed by a consumer operation unit of the dataflow graph;
• identifying a group of consumer operation units having a same read access pattern, comparing the write access pattern to the read access pattern;
• and
• classifying a data access pattern between the producer operation unit and the group of consumer operation units as streaming compatible if the read access pattern is same as the write access pattern,
• classifying the data access pattern between the producer operation unit and the group of consumer operation units as broadcast compatible if the read access pattern is different from the write access pattern, and
• modifying the dataflow graph by inserting only a single buffer between the producer operation unit and the group of consumer operation units only if the data access pattern is broadcast compatible to create a modified dataflow graph, thereby resolving the difference between the write access pattern and the read access pattern using only the single buffer.